Merge pull request #8668 from KonduitAI/master

Update master - recent fixes

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/CapsnetGradientCheckTest.java

@@ -43,12 +43,6 @@ import java.util.Random;
 
 public class CapsnetGradientCheckTest extends BaseDL4JTest {
 
-    private static final boolean PRINT_RESULTS = true;
-    private static final boolean RETURN_ON_FIRST_FAILURE = false;
-    private static final double DEFAULT_EPS = 1e-6;
-    private static final double DEFAULT_MAX_REL_ERROR = 1e-3;
-    private static final double DEFAULT_MIN_ABS_ERROR = 1e-8;
-
     @Test
     public void testCapsNet() {
 

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/OutputLayerGradientChecks.java

@@ -43,10 +43,6 @@ import static org.junit.Assert.assertTrue;
 public class OutputLayerGradientChecks extends BaseDL4JTest {
 
     private static final boolean PRINT_RESULTS = true;
-    private static final boolean RETURN_ON_FIRST_FAILURE = false;
-    private static final double DEFAULT_EPS = 1e-6;
-    private static final double DEFAULT_MAX_REL_ERROR = 1e-3;
-    private static final double DEFAULT_MIN_ABS_ERROR = 1e-8;
 
     static {
         Nd4j.setDataType(DataType.DOUBLE);

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/RnnGradientChecks.java

@@ -47,10 +47,6 @@ import static org.junit.Assert.assertTrue;
 public class RnnGradientChecks extends BaseDL4JTest {
 
     private static final boolean PRINT_RESULTS = true;
-    private static final boolean RETURN_ON_FIRST_FAILURE = false;
-    private static final double DEFAULT_EPS = 1e-6;
-    private static final double DEFAULT_MAX_REL_ERROR = 1e-3;
-    private static final double DEFAULT_MIN_ABS_ERROR = 1e-8;
 
     static {
         Nd4j.setDataType(DataType.DOUBLE);

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/UtilLayerGradientChecks.java

@@ -48,12 +48,6 @@ import static org.junit.Assert.assertTrue;
 
 public class UtilLayerGradientChecks extends BaseDL4JTest {
 
-    private static final boolean PRINT_RESULTS = true;
-    private static final boolean RETURN_ON_FIRST_FAILURE = false;
-    private static final double DEFAULT_EPS = 1e-6;
-    private static final double DEFAULT_MAX_REL_ERROR = 1e-3;
-    private static final double DEFAULT_MIN_ABS_ERROR = 1e-6;
-
     static {
         Nd4j.setDataType(DataType.DOUBLE);
     }
@@ -182,9 +176,9 @@ public class UtilLayerGradientChecks extends BaseDL4JTest {
                     MultiLayerNetwork net = new MultiLayerNetwork(conf);
                     net.init();
 
-                    boolean gradOK = GradientCheckUtil.checkGradients(new GradientCheckUtil.MLNConfig().net(net).input(input)
+                    boolean gradOK = GradientCheckUtil.checkGradients(new GradientCheckUtil.MLNConfig().net(net)
-                            .minAbsoluteError(1e-7)
+                            .minAbsoluteError(1e-6)
-                            .labels(label).inputMask(inMask));
+                            .input(input).labels(label).inputMask(inMask));
                     assertTrue(gradOK);
 
                     TestUtils.testModelSerialization(net);
@@ -233,8 +227,9 @@ public class UtilLayerGradientChecks extends BaseDL4JTest {
             //Test ComputationGraph equivalent:
             ComputationGraph g = net.toComputationGraph();
 
-            boolean gradOKCG = GradientCheckUtil.checkGradients(new GradientCheckUtil.GraphConfig().net(g).inputs(new INDArray[]{in})
+            boolean gradOKCG = GradientCheckUtil.checkGradients(new GradientCheckUtil.GraphConfig().net(g)
-                    .labels(new INDArray[]{labels}).excludeParams(excludeParams));
+                    .minAbsoluteError(1e-6)
+                    .inputs(new INDArray[]{in}).labels(new INDArray[]{labels}).excludeParams(excludeParams));
             assertTrue(gradOKCG);
 
             TestUtils.testModelSerialization(g);

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/YoloGradientCheckTests.java

@@ -56,11 +56,6 @@ import static org.junit.Assert.assertTrue;
  * @author Alex Black
  */
 public class YoloGradientCheckTests extends BaseDL4JTest {
-    private static final boolean PRINT_RESULTS = true;
-    private static final boolean RETURN_ON_FIRST_FAILURE = false;
-    private static final double DEFAULT_EPS = 1e-6;
-    private static final double DEFAULT_MAX_REL_ERROR = 1e-3;
-    private static final double DEFAULT_MIN_ABS_ERROR = 1e-8;
 
     static {
         Nd4j.setDataType(DataType.DOUBLE);

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/pooling/GlobalPoolingMaskingTests.java

@@ -21,14 +21,13 @@ import org.deeplearning4j.nn.conf.ConvolutionMode;
 import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
 import org.deeplearning4j.nn.conf.distribution.NormalDistribution;
-import org.deeplearning4j.nn.conf.layers.ConvolutionLayer;
+import org.deeplearning4j.nn.conf.layers.*;
-import org.deeplearning4j.nn.conf.layers.GravesLSTM;
+import org.deeplearning4j.nn.conf.layers.GlobalPoolingLayer;
-import org.deeplearning4j.nn.conf.layers.OutputLayer;
-import org.deeplearning4j.nn.conf.layers.PoolingType;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
 import org.deeplearning4j.nn.weights.WeightInit;
 import org.junit.Test;
 import org.nd4j.linalg.activations.Activation;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.ops.impl.broadcast.BroadcastMulOp;
 import org.nd4j.linalg.factory.Nd4j;
@@ -416,4 +415,53 @@ public class GlobalPoolingMaskingTests extends BaseDL4JTest {
             }
         }
     }
+
+    @Test
+    public void testMaskLayerDataTypes(){
+
+        for(DataType dt : new DataType[]{DataType.FLOAT16, DataType.BFLOAT16, DataType.FLOAT, DataType.DOUBLE,
+                DataType.INT8, DataType.INT16, DataType.INT32, DataType.INT64,
+                DataType.UINT8, DataType.UINT16, DataType.UINT32, DataType.UINT64}){
+            INDArray mask = Nd4j.rand(DataType.FLOAT, 2, 10).addi(0.3).castTo(dt);
+
+            for(DataType networkDtype : new DataType[]{DataType.FLOAT16, DataType.BFLOAT16, DataType.FLOAT, DataType.DOUBLE}){
+
+                INDArray in = Nd4j.rand(networkDtype, 2, 5, 10);
+                INDArray label1 = Nd4j.rand(networkDtype, 2, 5);
+                INDArray label2 = Nd4j.rand(networkDtype, 2, 5, 10);
+
+                for(PoolingType pt : PoolingType.values()) {
+                    //System.out.println("Net: " + networkDtype + ", mask: " + dt + ", pt=" + pt);
+
+                    MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
+                            .list()
+                            .layer(new GlobalPoolingLayer(pt))
+                            .layer(new OutputLayer.Builder().nIn(5).nOut(5).activation(Activation.TANH).lossFunction(LossFunctions.LossFunction.MSE).build())
+                            .build();
+
+                    MultiLayerNetwork net = new MultiLayerNetwork(conf);
+                    net.init();
+
+                    net.output(in, false, mask, null);
+                    net.output(in, false, mask, null);
+
+
+                    MultiLayerConfiguration conf2 = new NeuralNetConfiguration.Builder()
+
+                            .list()
+                            .layer(new RnnOutputLayer.Builder().nIn(5).nOut(5).activation(Activation.TANH).lossFunction(LossFunctions.LossFunction.MSE).build())
+                            .build();
+
+                    MultiLayerNetwork net2 = new MultiLayerNetwork(conf2);
+                    net2.init();
+
+                    net2.output(in, false, mask, mask);
+                    net2.output(in, false, mask, mask);
+
+                    net.fit(in, label1, mask, null);
+                    net2.fit(in, label2, mask, mask);
+                }
+            }
+        }
+    }
 }

deeplearning4j/deeplearning4j-core/src/test/resources/logback-test.xml

@@ -19,8 +19,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-cuda/src/test/resources/logback.xml

@@ -19,8 +19,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/src/test/resources/logback.xml

@@ -19,8 +19,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/RnnOutputLayer.java

@@ -131,7 +131,7 @@ public class RnnOutputLayer extends BaseOutputLayer<org.deeplearning4j.nn.conf.l
         INDArray W = getParamWithNoise(DefaultParamInitializer.WEIGHT_KEY, training, workspaceMgr);
 
         applyDropOutIfNecessary(training, workspaceMgr);
-        INDArray input2d = TimeSeriesUtils.reshape3dTo2d(input, LayerWorkspaceMgr.noWorkspaces(), ArrayType.FF_WORKING_MEM);
+        INDArray input2d = TimeSeriesUtils.reshape3dTo2d(input.castTo(W.dataType()), workspaceMgr, ArrayType.FF_WORKING_MEM);
 
         INDArray act2d = layerConf().getActivationFn().getActivation(input2d.mmul(W).addiRowVector(b), training);
         if (maskArray != null) {

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/util/MaskedReductionUtil.java

@@ -56,6 +56,7 @@ public class MaskedReductionUtil {
             throw new IllegalArgumentException("Expect rank 2 array for mask: got " + mask.rank());
         }
 
+        toReduce = toReduce.castTo(dataType);
         mask = mask.castTo(dataType);
 
         //Sum pooling: easy. Multiply by mask, then sum as normal
@@ -64,13 +65,7 @@ public class MaskedReductionUtil {
 
         switch (poolingType) {
             case MAX:
-                //TODO This is ugly - replace it with something better... Need something like a Broadcast CAS op
+                INDArray negInfMask = mask.castTo(dataType).rsub(1.0);
-                INDArray negInfMask;
-                if(mask.dataType() == DataType.BOOL){
-                    negInfMask = Transforms.not(mask).castTo(dataType);
-                } else {
-                    negInfMask = mask.rsub(1.0);
-                }
                 BooleanIndexing.replaceWhere(negInfMask, Double.NEGATIVE_INFINITY, Conditions.equals(1.0));
 
                 INDArray withInf = Nd4j.createUninitialized(dataType, toReduce.shape());
@@ -121,18 +116,14 @@ public class MaskedReductionUtil {
         //Mask: [minibatch, tsLength]
         //Epsilon: [minibatch, vectorSize]
 
+        mask = mask.castTo(input.dataType());
+
         switch (poolingType) {
             case MAX:
-                //TODO This is ugly - replace it with something better... Need something like a Broadcast CAS op
+                INDArray negInfMask = mask.rsub(1.0);
-                INDArray negInfMask;
-                if(mask.dataType() == DataType.BOOL){
-                    negInfMask = Transforms.not(mask).castTo(Nd4j.defaultFloatingPointType());
-                } else {
-                    negInfMask = mask.rsub(1.0);
-                }
                 BooleanIndexing.replaceWhere(negInfMask, Double.NEGATIVE_INFINITY, Conditions.equals(1.0));
 
-                INDArray withInf = Nd4j.createUninitialized(input.shape());
+                INDArray withInf = Nd4j.createUninitialized(input.dataType(), input.shape());
                 Nd4j.getExecutioner().exec(new BroadcastAddOp(input, negInfMask, withInf, 0, 2));
                 //At this point: all the masked out steps have value -inf, hence can't be the output of the MAX op
 
@@ -145,7 +136,7 @@ public class MaskedReductionUtil {
                 //if out = avg(in,dims) then dL/dIn = 1/N * dL/dOut
                 //With masking: N differs for different time series
 
-                INDArray out = Nd4j.createUninitialized(input.shape(), 'f');
+                INDArray out = Nd4j.createUninitialized(input.dataType(), input.shape(), 'f');
 
                 //Broadcast copy op, then divide and mask to 0 as appropriate
                 Nd4j.getExecutioner().exec(new BroadcastCopyOp(out, epsilon2d, out, 0, 1));
@@ -162,7 +153,7 @@ public class MaskedReductionUtil {
 
             case PNORM:
                 //Similar to average and sum pooling: there's no N term here, so we can just set the masked values to 0
-                INDArray masked2 = Nd4j.createUninitialized(input.shape());
+                INDArray masked2 = Nd4j.createUninitialized(input.dataType(), input.shape());
                 Nd4j.getExecutioner().exec(new BroadcastMulOp(input, mask, masked2, 0, 2));
 
                 INDArray abs = Transforms.abs(masked2, true);

deeplearning4j/deeplearning4j-remote/deeplearning4j-json-server/src/test/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/src/test/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/src/test/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/test/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/test/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-model/src/test/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-standalone/src/main/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui/src/test/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-vertx/src/test/resources/logback.xml

@@ -21,8 +21,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

deeplearning4j/dl4j-integration-tests/src/test/resources/logback-test.xml

@@ -19,8 +19,8 @@
     <appender name="FILE" class="ch.qos.logback.core.FileAppender">
         <file>logs/application.log</file>
         <encoder>
-            <pattern>%date - [%level] - from %logger in %thread
+            <pattern> %logger{15} - %message%n%xException{5}
-                %n%message%n%xException%n</pattern>
+            </pattern>
         </encoder>
     </appender>
 

libnd4j/CMakeLists.txt

@@ -5,7 +5,7 @@ option(NATIVE "Optimize for build machine (might not work on others)" OFF)
 set(CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/cmake" ${CMAKE_MODULE_PATH})
 #ensure we create lib files
 set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS OFF)
-
+option(CHECK_VECTORIZATION "checks for vectorization" OFF)
 option(BUILD_TESTS "Build tests" OFF)
 option(FLATBUFFERS_BUILD_FLATC "Enable the build of the flatbuffers compiler" OFF)
 set(FLATBUFFERS_BUILD_FLATC "OFF" CACHE STRING "Hack to disable flatc build" FORCE)

libnd4j/CMakeLists.txt.mkldnn.in

@@ -5,7 +5,7 @@ project(mkldnn-download NONE)
 include(ExternalProject)
 ExternalProject_Add(mkldnn
   GIT_REPOSITORY     https://github.com/intel/mkl-dnn.git
-  GIT_TAG           v1.1.2
+  GIT_TAG           v1.1.3
   SOURCE_DIR        "${CMAKE_CURRENT_BINARY_DIR}/mkldnn-src"
   BINARY_DIR        "${CMAKE_CURRENT_BINARY_DIR}/mkldnn-build"
   CONFIGURE_COMMAND ""

libnd4j/README.md

@@ -17,8 +17,11 @@ There's few additional arguments for `buildnativeoperations.sh` script you could
  -b release OR -b debug // enables/desables debug builds. release is considered by default
  -j XX // this argument defines how many threads will be used to binaries on your box. i.e. -j 8 
  -cc XX// CUDA-only argument, builds only binaries for target GPU architecture. use this for fast builds
+ --check-vectorization  auto-vectorization report for developers. (Currently, only GCC is supported)
 ```
 
+[More about AutoVectorization report](auto_vectorization/AutoVectorization.md)  
+
 You can find the compute capability for your card [on the NVIDIA website here](https://developer.nvidia.com/cuda-gpus).
 
 For example, a GTX 1080 has compute capability 6.1, for which you would use ```-cc 61``` (note no decimal point).

libnd4j/auto_vectorization/AutoVectorization.md

@@ -0,0 +1,49 @@
+# Auto-vectorization Report
+
+This report tool is used to get a human-friendly compiler output of the auto-vectorization process. It is intended for developers to help them to investigate the obstacles that compiler faced during auto-vectorization.
+
+## Usage
+```--check-vectorization``` option should be added to the **release** build to be able to get the auto-vectorization report 
+```./buildnativeoperations.sh -a native  -j 28 --check-vectorization```
+it will output ```vecmiss.html``` inside  blasbuild/cpu folder.
+
+## Report Format 
+Each filename contains info about optimization attempts for the source code lines. 
+Each line number is also expandable (⇲) and contains distinct failure notes. 
+It is possible to click on the line number to see source code  
+
+| file name  | total successful attempts | total failed attempts | ⇲  |
+|---|---|---|--|
+| line number  | successful attempts  | failed attempts  | ⇲ |
+|- failure reasons  |
+| line number  | successful attempts  | failed attempts  |⇲ |
+
+##### Requirements 
+- GCC (Currently, only GCC is supported)
+- python3
+
+### Detailed report with `-fsave-optimization-record` option:
+If you want to get more detailed information (for now it reports the functions of failures)  you should use new version of the toolchain (GCC > 9). As the new version of GCC compilers have `-fsave-optimization-record` option.   
+`buildnativeoperations.sh` using CMake will detect it and switch to the more detailed version.
+Please, note that this option is still experimental and so the compiler can fail to output some json.gz file with error.
+On that case try to exclude those files from the build.  
+And also the internal structure of the `-fsave-optimization-record` json.gz  can be changed in future. 
+
+It outputs two files **vecmiss_fsave.html** and **vecmiss_fsave.html.js**. So to see report details you need to enable javascript on browser if it was disabled.
+
+##### Requirements for the Detailed report
+- GCC version > 9
+- python3 
+- Cython (python3)
+- json   (python3)
+- gzip   (python3) 
+- c++filt
+
+Internally, we are using Cython to speed up json.gz file processing (bigGzipJson.pyx). Because json.gz files can take big memory in raw when loaded in whole.
+
+If you want to use bigGzipJson outside  `buildnativeoperations.sh` and CMake then you should compile it manually using this command in auto_vectorization folder:
+`python3 cython_setup.py build_ext --inplace`  
+
+json.gz files could be processed outside of `buildnativeoperations.sh`.    
+You need to call `python3 auto_vect.py --fsave` inside base source folder and where json.gz files exist.  
+

libnd4j/auto_vectorization/auto_vect.py

@@ -0,0 +1,546 @@
+'''
+@author : Abdelrauf rauf@konduit.ai
+'''
+import re
+import sys
+import os
+import subprocess
+import fnmatch
+import json
+import gzip
+try:
+    from bigGzipJson import json_gzip_extract_objects
+except ImportError:
+    pass
+from pathlib import Path
+from multiprocessing import  Pool, Manager ,cpu_count
+import traceback
+import html
+
+mtch = re.compile(r"[^/]*([^:]+)\:(\d+)\:(\d+)\:(.*)")
+replace_msg = re.compile(r"(\d+)?\.?(\d+)?_?\d+\.?(\d+)?")
+progress_msg = re.compile(r"\s{0,4}\[\s{0,2}\d+\%\]")
+file_dir_strip = str(Path(os.getcwd()))
+pp_index = file_dir_strip.rfind("libnd4j")
+if pp_index>=0:
+    file_dir_strip =file_dir_strip[:pp_index+len("libnd4j")]
+BASE_URL = "https://github.com/eclipse/deeplearning4j/tree/master/libnd4j/"
+if BASE_URL.endswith("/")==False:
+    BASE_URL = BASE_URL + "/"
+#print(file_dir_strip)
+class info:
+    def __repr__(self):
+        return str(self.__dict__) 
+
+FSAVE_IGNORE_EXTERNALS = True
+
+def get_cxx_filt_result(strx):
+    if len(strx)<1:
+        return ""
+    res = subprocess.Popen(["c++filt","-i", strx], stdout=subprocess.PIPE).communicate()[0]
+    res =res.decode('utf-8')
+    #replace some long names to reduce size
+    res = res.replace("unsigned long long", "uLL")
+    res = res.replace("unsigned long int","uL")
+    res = res.replace("unsigned long", "uL")
+    res = res.replace("unsigned int", "ui")
+    res = res.replace("unsigned char", "uchar")
+    res = res.replace("unsigned short", "ushort")    
+    res = res.replace("long long", "LL")
+    res = res.replace(", ",",")
+    return res.strip()
+
+ 
+def internal_glob(dir, match):
+    listx = []
+    for root, dirnames, filenames in os.walk(dir):
+        for filename in fnmatch.filter(filenames, match):
+            listx.append(os.path.join(root, filename))
+    return listx
+
+def get_obj_json_gz(filename):
+    with gzip.GzipFile(filename, 'r') as f:
+        return json.loads(f.read().decode('utf-8'))[-1]
+
+
+
+def get_msg(msg):
+    msg = msg.lower().strip() 
+    if "note: not vectorized:" in msg:
+        msg = replace_msg.sub("_numb",msg.replace("note: not vectorized:",""))
+        return( 0, 1, msg.strip()) 
+    elif "loop vectorized" in msg: 
+        return (1, 0, None)
+    # elif msg.startswith("missed")==False:
+    #     msg = replace_msg.sub("_numb",msg)
+    #     return( 0, 0, msg.strip())         
+    return None
+
+
+ 
+
+class File_Info:
+    '''
+    Holds information about vectorized and miss vectorized lines for one file
+    '''
+    
+    def __init__(self):
+        self.infos = {}
+        self.total_opted =0
+        self.total_missed = 0
+        self.external = False
+
+
+    def add_line(self, line_pos):
+        if line_pos not in self.infos:
+            v = info()
+            v.optimized = 0
+            v.missed = 0
+            v.miss_details = set()
+            self.infos[line_pos] = v
+            return v
+        else:
+            return self.infos[line_pos]
+
+
+    def add_line_fsave(self, line_pos):
+        if line_pos not in self.infos:
+            v = info()
+            v.optimized = 0
+            v.missed = 0
+            v.miss_details2 = dict()
+            self.infos[line_pos] = v
+            return v
+        else:
+            return self.infos[line_pos]            
+               
+
+
+    def add_fsave(self, line_pos,success, msg, function ,inline_fns=''):
+        v = self.add_line_fsave(line_pos)
+        if success and "loop vectorized" in msg:
+            v.optimized +=1
+            self.total_opted +=1
+        elif success==False and "not vectorized:" in msg:
+            #reduce this msg
+            msg = msg.replace("not vectorized:","")
+            v.missed +=1
+            self.total_missed +=1
+            msg = sys.intern(msg)
+            if msg in v.miss_details2:
+                ls = v.miss_details2.get(msg) 
+                ls.add(function)
+            else:
+                ls =set()
+                v.miss_details2[msg]=ls
+                ls.add(function)
+        return self
+
+    def add(self, line_pos, msg_x):
+        v = self.add_line(line_pos)
+        if msg_x is not None:
+                v.optimized += msg_x[0]
+                v.missed += msg_x[1]
+                self.total_opted += msg_x[0]
+                self.total_missed += msg_x[1]
+                if msg_x[2] is not None:
+                    v.miss_details.add(msg_x[2])
+        return self
+    
+
+    def __repr__(self):
+        return str(self.__dict__)                    
+                    
+
+
+
+def process_gzip_json_mp(args):
+    process_gzip_json_new(*args)
+
+def process_gzip_json_new(json_gz_fname,list_Queue):
+    gz_name = Path(json_gz_fname).stem
+    #print("::--open and process {0}".format(gz_name))
+    queue_count = len(list_Queue)
+    #print(queue_count)
+    q = list_Queue[0]
+    old_fname = ''
+    total_c = 0
+    for x in json_gzip_extract_objects(json_gz_fname,'message','vectorized'):
+        external_source = True
+        if len(x['message'])>0 and 'location' in x:
+            line = int(x['location']['line'])
+            file_name = x['location']['file'].strip()
+            if  file_dir_strip in file_name:
+                file_name = file_name.replace(file_dir_strip,'./')
+                external_source = False
+            msg = x['message'][0]
+            success = x['kind'] == 'success'
+            func = ''  if 'function' not in x else x['function']
+
+            if file_name!=old_fname:
+                #send our info to the right consumer
+                queue_ind = hash(file_name) % queue_count
+                #print("quen index {0}".format(queue_ind))
+                q =list_Queue[queue_ind]
+                old_fname = file_name
+            total_c +=1
+            #print("pp {0} {1}".format(q,(file_name,line,success, msg, func,external_source )))
+            if FSAVE_IGNORE_EXTERNALS==True and external_source == True:
+                continue
+            q.put((file_name,line,success, msg, func,external_source ))
+    print("::finished {0:60s} :{1:8d}".format(gz_name,total_c))
+
+def consume_processed_mp(args):
+    return consume_processed_new(*args)
+
+
+
+def consume_processed_new(list_Queue , c_index):
+    
+    info_ = dict()
+    func_list = dict()
+    last_func_index = 0 
+    q  = list_Queue[c_index]
+    print("::consumer {0}".format(c_index))
+    total_c = 0
+    r_c = 0
+    while True:
+        #print("try to get new from {0}".format(index))
+        obj = q.get()
+        #print("cc {0} {1}".format(q,obj))
+        if obj==None:
+            break #we received the end
+        file_name,line,success, msg, func, external_source = obj
+        try:
+            #get function index
+            func_index = -1
+            if func in func_list:
+                func_index = func_list[func]
+            else:
+                func_list[func] = last_func_index
+                func_index = last_func_index
+                last_func_index +=1
+ 
+            if file_name in info_: 
+                info_[file_name].add_fsave(line, success, msg, func_index)
+            else: 
+                info_[file_name] = File_Info().add_fsave(line, success, msg, func_index)
+                info_[file_name].external = external_source
+            total_c +=1 
+            if total_c - r_c >10000:
+                r_c = total_c
+                print("::consumer {0:2d} :{1:10d}".format(c_index,total_c))
+        except Exception as e:
+            print(traceback.format_exc())
+            break
+
+    print("::consumer {0:2d} :{1:10d}".format(c_index,total_c))
+    #write to temp file 
+    wr_fname= "vecmiss_fsave{0}.html".format(str(c_index) if len(list_Queue)>1 else '')
+    print("generate report for consumer {0} {1}".format(c_index,len(info_)))
+    try:
+        uniq_ind = str(c_index)+'_' if len(list_Queue)>1  else ''
+        generate_report(wr_fname,info_ ,only_body = False, unique_id_prefix = uniq_ind,fsave_format = True, function_list= func_list)
+        print(" consumer {0} saved output into {1}".format(c_index,wr_fname))
+    except Exception as e:
+        print(traceback.format_exc())
+
+            
+
+def obtain_info_from(input_):
+    info_ = dict()
+    for line in input_:
+        x = mtch.match(line)
+        external_source = True
+        if x:
+            file_name =x.group(1).strip()
+            if  file_dir_strip in file_name:
+                file_name = file_name.replace(file_dir_strip,'')
+                external_source = False
+            line_number = int(x.group(2))
+            msg = x.group(4).lower()
+            msg = msg.replace(file_dir_strip,'./')
+            msg_x = get_msg(msg)
+            if msg_x is None:
+                continue
+            if file_name in info_:
+                #ignore col_number
+                info_[file_name].add(line_number,msg_x)
+            else:
+                #print("{0} {1}".format(file_name,external_source))
+                info_[file_name] = File_Info().add(line_number,msg_x)
+                info_[file_name].external = external_source
+        elif progress_msg.match(line):
+            #actually we redirect only, stderr so this should not happen
+            print("__"+line.strip())
+        elif "error" in line or "Error" in line:
+            print("****"+line.strip())
+    return info_
+
+
+           
+def custom_style(fsave):
+    st = '''<style>a{color:blue;}
+a:link{text-decoration:none}a:visited{text-decoration:none}a:hover{cursor:pointer;text-decoration:underline}
+a:active{text-decoration:underline}
+.f.ext{display:none} 
+.f{color:#000;display:flex;overflow:hidden;justify-content:space-between;flex-wrap:wrap;align-items:baseline;width:100%}
+.f>div{min-width:10%}.f>div:first-child{min-width:70%;text-overflow:ellipsis}
+.f:nth-of-type(even){background-color:#f5f5f5}
+.f>div.g{flex:0 0 100%}.f>div:nth-child(2){font-weight:600;color:green}
+.f>div:nth-child(3){font-weight:600;color:red}
+.f>div:nth-child(2)::after{content:' ✓';color:green}.f>div:nth-child(3)::after{content:' -';color:red}
+.f>div.g>div>div:nth-child(2){font-weight:600;color:green}
+.f>div.g>div>div:nth-child(3){font-weight:600;color:red}
+.f>div.g>div>div:nth-child(2)::after{content:' ✓';color:green}
+.f>div.g>div>div:nth-child(3)::after{content:' -';color:red}
+.f>div.g>div{display:flex;justify-content:space-between;flex-wrap:wrap;align-items:baseline}
+.f>div.g>div>div{min-width:10%;text-align:left}
+.g>div:nth-of-type(even){background-color:#ede6fa}
+.f>div.g>div>ul{flex:0 0 100%}input[type=checkbox]{opacity:0;display:none}label{cursor:pointer}
+.f>label{color:red}input[type=checkbox]~.g{display:none}input[type=checkbox]:checked~.g{display:block}
+input[type=checkbox]~ul{display:none}
+input[type=checkbox]:checked~ul{display:block}input[type=checkbox]+label::after{content:"⇲";display:block}
+input[type=checkbox]:checked+label::after{content:"⇱";display:block}
+
+'''
+    if fsave==True:
+        st+='''.modal{display:none;height:100%;background-color:#144F84;color:#fff;opacity:.93;left:0;position:fixed;top:0;width:100%}
+        .modal.open{display:flex;flex-direction:column}.modal__header{height:auto;font-size:large;padding:10px;background-color:#000;color:#fff}
+        .modal__footer{height:auto;font-size:medium;background-color:#000}
+        .modal__content{height:100%;display:flex;flex-direction:column;padding:20px;overflow-y:auto}
+        .modal_close{cursor:pointer;float:right}li{cursor:pointer}
+        '''
+    return st + '''</style>'''
+
+def header(fsave=False):
+    strx ='<!DOCTYPE html>\n<html>\n<head>\n<meta charset="UTF-8">\n<title>Auto-Vectorization</title>\n'
+    strx +='<base id="base_id" href="{0}" target="_blank" >'.format(BASE_URL)
+    strx +=custom_style(fsave)  
+    strx +='\n</head>\n<body>\n'
+    return strx
+
+def footer():
+    return '\n</body></html>'
+
+
+def get_compressed_indices(set_a):
+    a_len = len(set_a)
+    if a_len<=1:
+        if a_len<1:
+            return ''
+        return str(set_a)[1:-1]
+    #we sorted and only saved difference 
+    # 1,14,15,19 --> 1,13,1,4  10bytes=>8bytes
+    list_sorted = sorted(list(set_a))
+    last = list_sorted[0]
+    str_x = str(list_sorted[0])
+    for i in range(1,a_len):
+        str_x += ','+str(list_sorted[i]-last)
+        last = list_sorted[i]
+    return str_x
+
+
+    
+
+
+def get_content(k, v,  unique_id_prefix = '', fsave_format=False):
+    inner_str=''
+    content = ''
+    inc_id = 0
+    for fk,fv in sorted(v.infos.items()):
+        if fsave_format==True:
+            inner_str+='<div><div><a>{0}</a></div><div>{1}</div><div>{2}</div><input type="checkbox" id="c{3}{4}"><label for="c{3}{4}"></label><ul>'.format(
+            fk,fv.optimized,fv.missed,unique_id_prefix,inc_id)
+        else:    
+            inner_str+='<div><div><a href=".{0}#L{1}">{1}</a></div><div>{2}</div><div>{3}</div><input type="checkbox" id="c{4}{5}"><label for="c{4}{5}"></label><ul>'.format(
+            k,fk,fv.optimized,fv.missed,unique_id_prefix,inc_id)
+        inc_id+=1
+        if fsave_format==True:
+            #
+            for dt,df in fv.miss_details2.items():
+                #inner_str  +='<li data-fns="{0}">{1}</li>'.format(str(df).replace(", ",",")[1:-1],dt) 
+                inner_str  +='<li data-fns="{0}">{1}</li>'.format(get_compressed_indices(df),dt)           
+        else:
+            for dt in fv.miss_details:
+                inner_str+="<li>"+str(dt)+ "</li>"
+        inner_str+="</ul></div>\n"
+
+    content += '<div class="f'
+    if v.external:
+        content += " ext"
+    content += '">\n<div>{0}</div><div>{1}</div><div>{2}</div><input type="checkbox" id="i{3}{4}"><label for="i{3}{4}"></label>'.format(
+        k,v.total_opted,v.total_missed,unique_id_prefix,inc_id)  
+    content += "<div class='g'>"  
+    content += inner_str
+    content += "</div> </div>\n"
+    return content   
+
+
+def jscript_head():
+    return '''
+    window.onload = function () {
+    var modal = document.getElementsByClassName("modal")[0];
+    var modal_close = document.getElementsByClassName("modal_close")[0];
+    var content = document.getElementsByClassName("modal__content")[0];
+    a_tags = document.getElementsByTagName("a");
+    base_href = document.getElementById("base_id").href;
+    for(i=0;i<a_tags.length;i++){
+        a_tags[i].addEventListener("click", function () {
+            var source = event.target || event.srcElement;
+            file_src = source.parentElement.parentElement.parentElement.parentElement.children[0].innerText ;
+            link = base_href + file_src+'#L'+ source.innerText;
+            window.open(link, '_blank');
+            
+        });
+    }
+    modal_close.addEventListener("click", function () {
+        content.innerHTML = '';
+        modal.className = 'modal';
+    });
+    
+    ''' 
+def jscipt_end():
+    return '''
+    tags = document.getElementsByTagName("li");
+    function escapeHtml(unsafe) {
+    return unsafe
+         .replace(/&/g, "&amp;")
+         .replace(/</g, "&lt;")
+         .replace(/>/g, "&gt;")
+         .replace(/"/g, "&quot;")
+         .replace(/'/g, "&#039;");
+    }
+    for (i = 0; i < tags.length; i++) {
+        tags[i].addEventListener("click", function () {
+            var source = event.target || event.srcElement;
+            funcs = source.dataset.fns.split(",")
+            strx = ''
+            //we saved differences,not real indices
+            last_ind = 0;
+            for (j = 0; j < funcs.length; j++) {
+                ind  = last_ind + parseInt(funcs[j]);
+                strx += "<p>" + escapeHtml(func_list[ind]) + "</p>";
+                last_ind = ind;
+            }
+            if (strx.length > 0) {
+                content.innerHTML = strx;
+                modal.className = 'modal open';
+            }
+
+        });
+    }
+ 
+    };'''
+
+def additional_tags(fsave):
+    if fsave==False:
+        return ''
+    #
+    return '''<script type='text/javascript'> 
+    var script = document.createElement('script'); script.src =  window.location.href+".js" ;
+    document.head.appendChild(script); 
+    </script>
+    <div class="modal">
+        <div class="modal__header">Functions <span class="modal_close">X</span></div>
+        <div class="modal__content"></div>
+        <div class="modal__footer">========</div>
+    </div>
+    '''
+
+def generate_report(output_name,info_ ,only_body = False, unique_id_prefix='',fsave_format = False , function_list = None):
+    '''
+      Generate Auto-Vectorization Report in html format
+    '''
+
+    temp_str =''    
+    if fsave_format ==True:
+        # we gonna dump function_list as key list sorted by value
+        #and use it as jscript array  
+        sorted_funcs_by_index = sorted(function_list.items(), key=lambda x: x[1])
+        del function_list
+        with open(output_name+ ".js","w") as f:
+            #temp_str =jscript_head() +'{ "fmaps":['
+            temp_str = jscript_head() + "\n var func_list = ["
+            for k,v in sorted_funcs_by_index:
+                #json.dumps using for escape
+                #print(str(v)+str(k))
+                temp_str+=json.dumps(get_cxx_filt_result(k))+","
+                #reduce write calls
+                if len(temp_str)>8192*2:
+                    f.write(temp_str)
+                    temp_str= ''
+            if len(temp_str)>0:
+                f.write(temp_str)
+            f.write('"-"];'+jscipt_end())
+            
+
+    temp_str = ''
+    with open(output_name,"w") as f:
+        if only_body==False:
+            f.write(header(fsave_format))
+            f.write(additional_tags(fsave_format))
+        nm=0
+        for k,v in sorted(info_.items()): # sorted(info_.items(), key=lambda x: x[1].total_opted, reverse=True):
+            temp_str += get_content(k,v,unique_id_prefix+str(nm),fsave_format)
+            #reduce io write calls
+            if len(temp_str)>8192:
+                f.write(temp_str)
+                temp_str =''
+            nm+=1
+        if len(temp_str)>0:
+            f.write(temp_str)
+        if only_body==False:
+            f.write(footer())           
+
+
+def fsave_report_launch(json_gz_list):
+
+        cpus = cpu_count()
+        if cpus>32:
+            cpus = 24
+
+        c_count = 1 # 2 i sufficient # if cpus<=1 else min(4,cpus)
+        p_count = 3 if cpus<=1 else max(8, cpus - c_count)
+
+        m = Manager()
+        #consumer Queues
+        list_Queue = [m.Queue() for index in range(0,c_count)]
+        with Pool(processes=c_count) as consumers:
+            #start consumers
+            cs = consumers.map_async(consume_processed_mp,[(list_Queue, index,) for index in range(0,c_count)])
+            with Pool(processes=p_count) as processors:
+                processors.map(process_gzip_json_mp, [(fname, list_Queue,) for fname in json_gz_list])
+                
+            #send ends to inform our consumers
+            #send ends
+            for q in list_Queue:
+                q.put(None)
+            
+            #wait for consumers
+            cs.wait()
+ 
+
+
+
+def main():
+    if "--fsave" in sys.argv:
+        json_gz_list = internal_glob(".","*.json.gz")
+        fsave_report_launch(json_gz_list)
+        return        
+    
+    file_info = obtain_info_from(sys.stdin)
+    if len(file_info)>0:
+        #print(file_info)
+        print("---generating vectorization html report--")
+        generate_report("vecmiss.html", file_info)
+    else:
+        # lets check if we got fsave files
+        json_gz_list = internal_glob(".","*.json.gz")
+        fsave_report_launch(json_gz_list)
+
+
+
+
+if __name__ == '__main__':
+    main()     

libnd4j/auto_vectorization/bigGzipJson.pyx

@@ -0,0 +1,354 @@
+'''
+@author : Abdelrauf rauf@konduit.ai
+Simple object xtractor form very big json files 
+'''
+
+import sys
+from cpython.mem cimport PyMem_Malloc, PyMem_Realloc, PyMem_Free
+
+
+cdef char JSON_1 = b':'
+cdef char JSON_2 = b','
+cdef char JSON_3 = b'{'
+cdef char JSON_4 = b'}'
+cdef char JSON_5 = b'['
+cdef char JSON_6 = b']'
+cdef char QUOTE = b'"'
+cdef char ESCAPE = b"\\"
+cdef char SPACE = b' '
+cdef char TAB = b't'
+cdef char CR = b'\r'
+cdef char NL = b'\n'
+cdef char B = b'\b'
+cdef char EMPTY = b'\0'
+
+
+cdef struct Span:
+    int b
+    int e 
+
+cdef inline Span read_unquoted(char *text, int start,int end):
+    cdef Span sp
+    cdef int j = start
+    while j < end:
+        #if text[j].isspace():
+        if text[j] == SPACE or text[j] == NL or text[j] == TAB or text[j] == CR or text[j] == B:
+            j += 1
+            continue
+        if text[j] != QUOTE and text[j] != JSON_1 and text[j] != JSON_2 and text[j] != JSON_3 and text[j] != JSON_4 and text[j] != JSON_5 and text[j] != JSON_6:
+            start = j
+            j += 1
+            while j < end:
+                # read till JSON or white space
+                if text[j] == SPACE or text[j] == NL or text[j] == TAB or text[j] == CR or text[j] == B:
+                    sp.b = start
+                    sp.e = j
+                    return sp
+                elif text[j] == JSON_1 or text[j] == JSON_2 or text[j] == JSON_3 or text[j] == JSON_4 or text[j] == JSON_5 or text[j] == JSON_6:
+                    sp.b = start
+                    sp.e = j                 
+                    return sp
+                j += 1
+            if j == end-1:
+                sp.b = start
+                sp.e = end            
+                return sp
+        break
+    sp.b = j
+    sp.e = j
+    return sp
+
+
+cdef inline Span read_seq_token(char *text,int  start,int end):
+    #read quoted
+    #skip white_space
+    cdef Span sp    
+    cdef int j = start
+    cdef char last_char
+    cdef char char_x
+    while j < end:
+        if text[j] == SPACE or text[j] == NL or text[j] == TAB or text[j] == CR or text[j] == B:
+            j += 1
+            continue
+        if text[j] == QUOTE:
+            last_char = EMPTY
+            #read till another quote
+            start = j
+            j += 1
+            while j < end:
+                char_x = text[j]
+                if char_x == QUOTE and last_char != ESCAPE:
+                    # finished reading
+                    sp.b =start
+                    sp.e = j+1
+                    return sp
+                last_char = char_x
+                j += 1
+            if j == end-1:
+                sp.b = start
+                sp.e = end
+                return sp
+        else:
+            break
+    return read_unquoted(text, j, end)
+
+
+def tokenizer_spans(utext):
+    '''
+    we will just return tokenize spans
+    '''
+    token_spans = []
+    last_char = b''
+    end_i = len(utext)
+    cdef char *text = utext
+    i = 0
+    cdef Span sp
+    while i < end_i:
+        sp = read_seq_token(text, i, end_i)
+        i = sp.e
+        if sp.e > sp.b:
+            token_spans.append((sp.b, sp.e))
+        if i < end_i:
+            #if text[i] in JSON:
+            if text[i] == JSON_3 or text[i] == JSON_4 or text[i] == JSON_5 or text[i] == JSON_6 or text[i] == JSON_1 or text[i] == JSON_2:
+                token_spans.append((i, i+1))
+        i += 1
+    return token_spans
+
+
+
+
+
+cdef class JsonObjXtractor:
+    '''
+     JsonObjXtractor that utilize cython better
+    '''
+
+    cdef Span* token_spans
+    cdef size_t size      
+
+    def __cinit__(self, size_t count=4096): 
+        self.token_spans = <Span*> PyMem_Malloc(count * sizeof(Span))
+        self.size = count
+        if not self.token_spans:
+            raise MemoryError()
+
+
+    def __tokenizer_spans(self,utext, length):
+        '''
+        we will just return token  spans length
+        '''
+         
+        last_char = b''
+        end_i = length
+        cdef char *text = utext
+        cdef int i = 0
+        cdef size_t j = 0
+        cdef Span sp
+        while i < end_i:
+            sp = read_seq_token(text, i, end_i)
+            i = sp.e
+            if sp.e > sp.b:
+                self.token_spans[j] = sp
+                j+=1
+                if j>self.size:
+                    #we need to reallocate
+                    self.__resize(self.size+self.size//2)                
+            if i < end_i:
+                #if text[i] in JSON:
+                if text[i] == JSON_3 or text[i] == JSON_4 or text[i] == JSON_5 or text[i] == JSON_6 or text[i] == JSON_1 or text[i] == JSON_2:
+                    sp.b=i
+                    sp.e=i+1
+                    self.token_spans[j] = sp
+                    j+=1
+                    if j>self.size:
+                        #we need to reallocate
+                        self.__resize(self.size+self.size//2)
+            i += 1
+        return j
+
+
+
+    def try_extract_parent_obj(self, json_bytes, property_name, next_contains_value=b'', debug=False):
+        '''
+        try_extract_parent_obj(json_text, property_name, next_contains_value='', debug=False):
+        make sure that passed variables encoded to bytes with encode('utf-8')
+        next_contains_value either direct content or followed by '['
+        tries to extract the parent object for given named object
+        if the left brace of the parent object is outside of the current buffer
+        it will be ignored
+        if the right brace is outside of the buffer it will be left to be handled by caller
+        '''
+
+        look_for_the_left = True
+        parent_left = []
+        parent_right = []
+        parent_objects = []
+        len_next = len(next_contains_value) 
+        cdef int ind = 0
+        cdef int end
+        cdef int last_start = 0
+        property_name = b'"'+property_name+b'"'
+        cdef int lenx  = self.__tokenizer_spans(json_bytes,len(json_bytes))
+        cdef char x
+        cdef int i = -1
+        cdef Span sp
+        while i < lenx-1:
+            i += 1
+            ind = self.token_spans[i].b
+            x = json_bytes[ind]
+            #print("-----{0} -- {1} -- {2} ".format(x,parent_left,parent_right))
+            if look_for_the_left == False:
+                if x == JSON_3:
+                    parent_right.append(ind)
+                elif x == JSON_4:
+                    if len(parent_right) == 0:
+                        #we found parent closing brace
+                        look_for_the_left = True
+                        parent_objects.append((parent_left[-1], ind+1))
+                        last_start = ind+1
+                        #print("=============found {0}".format(parent_objects))
+                        parent_left = []
+                        parent_right = []
+                    else:
+                        parent_right.pop()
+                continue
+            #search obj
+            if look_for_the_left:
+                if x == JSON_3:
+                    parent_left.append(ind)
+                    last_start = ind
+                elif x == JSON_4:
+                    if len(parent_left) >= 1:
+                        #ignore
+                        parent_left.pop()
+
+            if x == JSON_1:  # ':'
+                #check to see if propertyname
+                old_property = EMPTY
+                if i > 1:
+                    sp = self.token_spans[i-1]
+                    old_property = json_bytes[sp.b:sp.e]
+                if old_property == property_name:
+                    #we found
+                    if len(parent_left) < 1:
+                        #left brace is outside of the buffer
+                        #we have to ignore it
+                        #try to increase buffer
+                        if debug:
+                            print('''left brace of the parent is outside of the buffer and parent is big.
+                            it will be ignored 
+                            try to choose disambiguous property names if you are looking for small objects''', file=sys.stderr)
+                        last_start = ind+1
+                        parent_left = []
+                        parent_right = []
+                        continue
+                    else:
+                        #print("++++++ look for the right brace")
+                        if len_next>0 and i+1 < lenx:
+                                i += 1
+                                ind = self.token_spans[i].b
+                                end = self.token_spans[i].e
+                                m = json_bytes[ind]
+                                
+                                if m == JSON_5:
+                                    #print ("----{0} {1}".format(m,JSON_5))
+                                    if i+1 < lenx:
+                                        i += 1
+                                        ind = self.token_spans[i].b
+                                        end = self.token_spans[i].e
+                                        #print ("----{0}  == {1}".format(next_contains_value,json_bytes[ind:end]))
+                                        if len_next <= end-ind and next_contains_value in json_bytes[ind:end]:
+                                            look_for_the_left = False
+                                            continue
+                                elif len_next <= end-ind and  next_contains_value in json_bytes[ind:end]:
+                                    look_for_the_left = False
+                                    continue
+
+                                #ignore as it does not have that value
+                                parent_left = []
+                                parent_right = []
+                                last_start = ind + 1
+                        else:
+                            look_for_the_left = False
+
+        # lets return last succesful opened brace as the last
+        # or  left brace failure case, safe closed brace
+        if len(parent_left)>0:
+            return (parent_objects, parent_left[-1])
+
+        return (parent_objects, last_start)
+
+
+
+    def __resize(self, size_t new_count):
+        cdef Span* mem = <Span*> PyMem_Realloc(self.token_spans, new_count * sizeof(Span))
+        if not mem:
+            raise MemoryError()
+        self.token_spans = mem
+        self.size = new_count
+
+    def __dealloc__(self):
+        PyMem_Free(self.token_spans) 
+
+
+
+import json
+import gzip
+import sys
+DEBUG_LOG = False
+
+def json_gzip_extract_objects(filename, property_name, next_contains_value=''):
+    strx = b''
+    started= False
+    b_next_contains_value = next_contains_value.encode('utf-8')
+    b_property_name = property_name.encode('utf-8')
+    #print(b_property_name)
+    objXt = JsonObjXtractor()
+    with gzip.open(filename, 'rb') as f:
+        if DEBUG_LOG:
+            print("opened {0}".format(filename), file=sys.stderr)
+        #instead of reading it as line, I will read it as binary bytes
+        is_End = False
+        #total = 0
+        while is_End==False:
+            buffer  = f.read(8192*2)
+            
+            lenx= len(buffer)
+            #total +=lenx
+            if lenx<1:
+                is_End = True
+            else:   
+                strx = strx + buffer
+
+            objects , last_index = objXt.try_extract_parent_obj(strx,b_property_name,b_next_contains_value)
+
+            # if b_property_name in strx and b_next_contains_value in strx:
+            #     print(strx)
+            #     print(objects)
+            #     print(last_index)
+            #     print("===================================================")
+
+            for start,end in objects:
+                yield  json.loads(strx[start:end]) #.decode('utf-8'))
+
+
+            #remove processed 
+            if last_index< len(strx):
+                strx = strx[last_index:]
+                
+            else:
+                strx = b''
+                #print('----+++')
+                
+            if(len(strx)>16384*3):
+                #buffer to big
+                #try to avoid big parents
+                if DEBUG_LOG:
+                    print("parent object is too big. please, look for better property name", file=sys.stderr)
+
+                break 
+
+
+
+

libnd4j/auto_vectorization/cython_setup.py

@@ -0,0 +1,3 @@
+from distutils.core import setup
+from Cython.Build import cythonize
+setup(ext_modules=cythonize("bigGzipJson.pyx",  language_level="3"))

libnd4j/blas/CMakeLists.txt

@@ -282,6 +282,32 @@ elseif(CPU_BLAS)
         set_source_files_properties(../include/helpers/impl/OpTracker.cpp PROPERTIES COMPILE_FLAGS "-march=x86-64 -mtune=generic")
     endif()
 
+    if(CHECK_VECTORIZATION)
+       set(VECT_FILES cpu/NativeOps.cpp ${OPS_SOURCES} ${HELPERS_SOURCES} ${CUSTOMOPS_GENERIC_SOURCES} ${LOOPS_SOURCES})
+       if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
+        
+        if (CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 9.0)
+            set(CHECK_VECT_FLAGS "-ftree-vectorize -fsave-optimization-record")
+            #to process fsave-optimization-record we will need our cython version code
+            message("Build Auto vectorization helpers")
+            execute_process(COMMAND "python3" "${CMAKE_CURRENT_SOURCE_DIR}/../auto_vectorization/cython_setup.py" "build_ext" "--inplace" WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/../auto_vectorization/" RESULT_VARIABLE ret)
+            message("build='${ret}'")
+
+            #remove fail cases that gcc fails produce sometimes
+            file(GLOB_RECURSE FAILURE_CASES false ../include/loops/cpu/compilation_units/reduce3*.cpp)
+            #message("*****${FAILURE_CASES}")
+            foreach(FL_ITEM ${FAILURE_CASES})
+                message("Removing failure cases ${FL_ITEM}")
+                list(REMOVE_ITEM VECT_FILES ${FL_ITEM})
+            endforeach() 
+        else()
+            set(CHECK_VECT_FLAGS "-ftree-vectorize -fopt-info-vec-optimized-missed")
+        endif()
+        message("CHECK VECTORIZATION ${CHECK_VECT_FLAGS}")
+        set_source_files_properties( ${VECT_FILES}  PROPERTIES COMPILE_FLAGS "${CHECK_VECT_FLAGS}" )
+       endif()
+    endif()    
+
     message("CPU BLAS")
     add_definitions(-D__CPUBLAS__=true)
     add_library(nd4jobj OBJECT cpu/NativeOps.cpp cpu/GraphExecutioner.cpp

libnd4j/blas/NDArray.h

@@ -195,6 +195,56 @@ namespace nd4j {
 
         NDArray(std::shared_ptr<DataBuffer> buffer, const char order, const std::vector<Nd4jLong> &shape, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
 
+        /**
+         * This contructors create scalar array containing string utf8
+         *
+         */
+        NDArray(const char* str, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext())
+            : NDArray(std::string(str), dtype, context) {
+        }
+        NDArray(const std::string& string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+
+        /**
+         * This contructors create scalar array containing string utf16
+         *
+         */
+        NDArray(const char16_t* u16string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext())
+            : NDArray(std::u16string(u16string), dtype, context) {
+        }
+
+        NDArray(const std::u16string& u16string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+
+        /**
+        * This contructors create scalar array containing string utf32
+        *
+        */
+        NDArray(const char32_t* u32string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext())
+            : NDArray(std::u32string(u32string), dtype, context) {
+        }
+
+        NDArray(const std::u32string& u32string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+
+        /**
+        * This contructors create array from vector of utf8 strings
+        *
+        */
+        NDArray(const std::vector<Nd4jLong>& shape, const std::vector<const char*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        NDArray(const std::vector<Nd4jLong>& shape, const std::vector<std::string>& string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+
+        /**
+         * This contructors create array from vector of utf16 strings
+         *
+        */
+        NDArray(const std::vector<Nd4jLong>& shape, const std::vector<const char16_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        NDArray(const std::vector<Nd4jLong>& shape, const std::vector<std::u16string>& string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+
+        /**
+         * This contructors create array from vector of utf32 strings
+         *
+        */
+        NDArray(const std::vector<Nd4jLong>& shape, const std::vector<const char32_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        NDArray(const std::vector<Nd4jLong>& shape, const std::vector<std::u32string>& string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+
 #endif
 
         /**
@@ -250,7 +300,6 @@ namespace nd4j {
         */
         NDArray(void *buffer, const char order, const std::vector<Nd4jLong> &shape,  nd4j::DataType dtype, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext(), const bool isBuffAlloc = false);
 
-
         /**
         * This method returns new array with the same shape & data type
         * @return
@@ -1148,6 +1197,9 @@ namespace nd4j {
         template <typename N>
         NDArray asT() const;
 
+        template <typename S>
+        NDArray asS() const;
+
         NDArray asT(DataType dtype) const;
 
 

libnd4j/blas/NDArrayFactory.h

@@ -1,5 +1,6 @@
 /*******************************************************************************
  * Copyright (c) 2015-2018 Skymind, Inc.
+ * Copyright (c) 2019-2020 Konduit K.K.
  *
  * This program and the accompanying materials are made available under the
  * terms of the Apache License, Version 2.0 which is available at
@@ -16,6 +17,7 @@
 
 //
 // Created by raver119 on 2018-09-16.
+// @author Oleg Semeniv <oleg.semeniv@gmail.com>
 //
 
 #ifndef DEV_TESTS_NDARRAYFACTORY_H
@@ -106,25 +108,72 @@ namespace nd4j {
         template <typename T>
         static NDArray create(char order, const std::vector<Nd4jLong> &shape, const std::initializer_list<T>& data, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
 
-        static NDArray string(const char *string, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        /**
+         * This factory create array from utf8 string
+         * @return NDArray default dataType UTF8
+         */
+        static NDArray string(const char *string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray* string_(const char *string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray* string_(const std::string &string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray string(const std::string& string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
 
-        static NDArray* string_(const char *string, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        /**
+         * This factory create array from utf16 string
+         * @return NDArray default dataType UTF16
+         */
+        static NDArray string(const char16_t* u16string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_(const char16_t* u16string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_(const std::u16string& u16string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray string(const std::u16string& u16string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
         
-        static NDArray string(const std::string &string, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        /**
+         * This factory create array from utf32 string
+         * @return NDArray default dataType UTF32
+         */
+        static NDArray string(const char32_t* u32string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_(const char32_t* u32string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_(const std::u32string& u32string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray string(const std::u32string& u32string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
 
-        static NDArray* string_(const std::string &string, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        /**
+         * This factory create array from vector of utf8 strings
+         * @return NDArray default dataType UTF8
+         */
+        static NDArray string( const std::vector<Nd4jLong> &shape, const std::initializer_list<const char *> &strings, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray string( const std::vector<Nd4jLong> &shape, const std::initializer_list<std::string> &string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray string( const std::vector<Nd4jLong> &shape, const std::vector<const char *> &strings, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray string( const std::vector<Nd4jLong> &shape, const std::vector<std::string> &string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong> &shape, const std::initializer_list<const char *> &strings, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong> &shape, const std::initializer_list<std::string> &string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong> &shape, const std::vector<const char *> &strings, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong> &shape, const std::vector<std::string> &string, nd4j::DataType dtype = nd4j::DataType::UTF8, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
 
-        static NDArray string(char order, const std::vector<Nd4jLong> &shape, const std::initializer_list<const char *> &strings, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        /**
-        static NDArray string(char order, const std::vector<Nd4jLong> &shape, const std::initializer_list<std::string> &string, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+         * This factory create array from vector of utf16 strings
+         * @return NDArray default dataType UTF16
+         */
+        static NDArray string( const std::vector<Nd4jLong>& shape, const std::initializer_list<const char16_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray string( const std::vector<Nd4jLong>& shape, const std::initializer_list<std::u16string>& string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray string( const std::vector<Nd4jLong>& shape, const std::vector<const char16_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray string( const std::vector<Nd4jLong>& shape, const std::vector<std::u16string>& string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong>& shape, const std::initializer_list<const char16_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong>& shape, const std::initializer_list<std::u16string>& string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong>& shape, const std::vector<const char16_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong>& shape, const std::vector<std::u16string>& string, nd4j::DataType dtype = nd4j::DataType::UTF16, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
 
-        static NDArray string(char order, const std::vector<Nd4jLong> &shape, const std::vector<const char *> &strings, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        /**
-        static NDArray string(char order, const std::vector<Nd4jLong> &shape, const std::vector<std::string> &string, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+         * This factory create array from vector of utf32 strings
+         * @return NDArray default dataType UTF32
+         */
+        static NDArray string( const std::vector<Nd4jLong>& shape, const std::initializer_list<const char32_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray string( const std::vector<Nd4jLong>& shape, const std::initializer_list<std::u32string>& string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray string( const std::vector<Nd4jLong>& shape, const std::vector<const char32_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray string( const std::vector<Nd4jLong>& shape, const std::vector<std::u32string>& string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong>& shape, const std::initializer_list<const char32_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong>& shape, const std::initializer_list<std::u32string>& string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong>& shape, const std::vector<const char32_t*>& strings, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
+        static NDArray* string_( const std::vector<Nd4jLong>& shape, const std::vector<std::u32string>& string, nd4j::DataType dtype = nd4j::DataType::UTF32, nd4j::LaunchContext* context = nd4j::LaunchContext::defaultContext());
 
-        static NDArray* string_(char order, const std::vector<Nd4jLong> &shape, const std::initializer_list<const char *> &strings, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
-        static NDArray* string_(char order, const std::vector<Nd4jLong> &shape, const std::initializer_list<std::string> &string, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
-
-        static NDArray* string_(char order, const std::vector<Nd4jLong> &shape, const std::vector<const char *> &strings, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
-        static NDArray* string_(char order, const std::vector<Nd4jLong> &shape, const std::vector<std::string> &string, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
 
         static ResultSet createSetOfArrs(const Nd4jLong numOfArrs, const void* buffer, const Nd4jLong* shapeInfo, const Nd4jLong* offsets, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
 

libnd4j/blas/NativeOps.h

@@ -1518,7 +1518,7 @@ ND4J_EXPORT int execCustomOp2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPoi
 typedef nd4j::ShapeList OpaqueShapeList;
 
 ND4J_EXPORT OpaqueShapeList* calculateOutputShapes(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs);
-ND4J_EXPORT OpaqueShapeList* calculateOutputShapes2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs);
+ND4J_EXPORT OpaqueShapeList* calculateOutputShapes2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs, int *dArgs, int numDArgs);
 
 ND4J_EXPORT Nd4jLong getShapeListSize(OpaqueShapeList* list);
 ND4J_EXPORT Nd4jLong* getShape(OpaqueShapeList* list, Nd4jLong i);
@@ -1607,6 +1607,7 @@ ND4J_EXPORT void setGraphContextInputArray(OpaqueContext* ptr, int index, void *
 ND4J_EXPORT void setGraphContextOutputArray(OpaqueContext* ptr, int index, void *buffer, void *shapeInfo, void *specialBuffer, void *specialShapeInfo);
 ND4J_EXPORT void setGraphContextInputBuffer(OpaqueContext* ptr, int index, OpaqueDataBuffer *buffer, void *shapeInfo, void *specialShapeInfo);
 ND4J_EXPORT void setGraphContextOutputBuffer(OpaqueContext* ptr, int index, OpaqueDataBuffer *buffer, void *shapeInfo, void *specialShapeInfo);
+ND4J_EXPORT void setGraphContextDArguments(OpaqueContext* ptr, int *arguments, int numberOfArguments);
 ND4J_EXPORT void setGraphContextTArguments(OpaqueContext* ptr, double *arguments, int numberOfArguments);
 ND4J_EXPORT void setGraphContextIArguments(OpaqueContext* ptr, Nd4jLong *arguments, int numberOfArguments);
 ND4J_EXPORT void setGraphContextBArguments(OpaqueContext* ptr, bool *arguments, int numberOfArguments);

libnd4j/blas/cpu/NDArrayFactory.cpp

@@ -1,5 +1,6 @@
 /*******************************************************************************
  * Copyright (c) 2015-2018 Skymind, Inc.
+ * Copyright (c) 2019-2020 Konduit K.K.
  *
  * This program and the accompanying materials are made available under the
  * terms of the Apache License, Version 2.0 which is available at
@@ -16,6 +17,7 @@
 
 //
 // Created by GS <sgazeos@gmail.com> on 2018-12-20.
+// @author Oleg Semeniv <oleg.semeniv@gmail.com>
 //
 
 #include <NDArrayFactory.h>
@@ -25,6 +27,9 @@
 #include <ShapeUtils.h>
 #include <type_traits>
 
+
+#include <StringUtils.h>
+
 namespace nd4j {
 
     ////////////////////////////////////////////////////////////////////////
@@ -85,45 +90,6 @@ namespace nd4j {
     template ND4J_EXPORT NDArray NDArrayFactory::create(const char order, const std::vector<Nd4jLong> &shape, const std::vector<uint8_t>& data, nd4j::LaunchContext * context);
     template ND4J_EXPORT NDArray NDArrayFactory::create(const char order, const std::vector<Nd4jLong> &shape, const std::vector<bool>& data, nd4j::LaunchContext * context);
 
-    NDArray NDArrayFactory::string(const char *str, nd4j::LaunchContext * context) {
-        std::string s(str);
-        return string(s, context);
-    }
-
-    NDArray* NDArrayFactory::string_(const char *str, nd4j::LaunchContext * context) {
-        return string_(std::string(str), context);
-    }
-
-    NDArray NDArrayFactory::string(const std::string &str, nd4j::LaunchContext * context) {
-
-        auto headerLength = ShapeUtils::stringBufferHeaderRequirements(1);
-
-        std::shared_ptr<DataBuffer> pBuffer = std::make_shared<DataBuffer>(headerLength + str.length(), DataType::UTF8, context->getWorkspace(), true);
-
-        NDArray res(pBuffer, ShapeDescriptor::scalarDescriptor(DataType::UTF8), context);
-
-        int8_t* buffer = reinterpret_cast<int8_t*>(res.getBuffer());
-
-        auto offsets = reinterpret_cast<Nd4jLong *>(buffer);
-        offsets[0] = 0;
-        offsets[1] = str.length();
-
-        auto data = buffer + headerLength;
-
-        memcpy(data, str.c_str(), str.length());
-
-        res.tickWriteHost();
-        res.syncToDevice();
-
-        return res;
-    }
-
-    NDArray* NDArrayFactory::string_(const std::string &str, nd4j::LaunchContext * context) {
-        auto res = new NDArray();
-        *res = NDArrayFactory::string(str, context);
-        return res;
-    }
-
 ////////////////////////////////////////////////////////////////////////
 template<typename T>
 NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, nd4j::LaunchContext * context) {
@@ -551,91 +517,175 @@ template ND4J_EXPORT NDArray NDArrayFactory::create(uint8_t * buffer, const char
 template ND4J_EXPORT NDArray NDArrayFactory::create(int8_t* buffer, const char order, const std::initializer_list<Nd4jLong>& shape, nd4j::LaunchContext * context);
 template ND4J_EXPORT NDArray NDArrayFactory::create(int16_t* buffer, const char order, const std::initializer_list<Nd4jLong>& shape, nd4j::LaunchContext * context);
 
+      /////////////////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string(const char16_t* u16string, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return NDArray(u16string, dtype, context); 
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_(const char16_t* u16string, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return string_(std::u16string(u16string), dtype, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_(const std::u16string& u16string, nd4j::DataType dtype, nd4j::LaunchContext* context) {  
+          auto res = new NDArray();
+          *res = NDArray(u16string, dtype, context);
+          return res;
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string(const std::u16string& u16string, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return NDArray(u16string, dtype, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string(const char32_t* u32string, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return NDArray(u32string, dtype, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_(const char32_t* u32string, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return string_(std::u32string(u32string), dtype, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_(const std::u32string& u32string, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          auto res = new NDArray();
+          *res = NDArray(u32string, dtype, context);
+          return res;
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string(const std::u32string& u32string, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return NDArray(u32string, dtype, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string(const char* str, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return NDArray(str, dtype, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_(const char* str, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return string_(std::string(str), dtype, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_(const std::string& str, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          auto res = new NDArray();
+          *res = NDArray(str, dtype, context);
+          return res;
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string(const std::string& str, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return NDArray(str, dtype, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string(const std::vector<Nd4jLong> &shape, const std::initializer_list<const char *> &strings, nd4j::DataType dataType, nd4j::LaunchContext * context) {
+          return NDArray(shape, std::vector<const char*>(strings), dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string( const std::vector<Nd4jLong> &shape, const std::vector<const char *> &strings, nd4j::DataType dataType, nd4j::LaunchContext * context) {
+          return NDArray( shape, strings, dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string( const std::vector<Nd4jLong> &shape, const std::initializer_list<std::string> &string, nd4j::DataType dataType, nd4j::LaunchContext * context) {
+          return NDArray( shape, std::vector<std::string>(string), dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong> &shape, const std::initializer_list<const char *> &strings, nd4j::DataType dataType, nd4j::LaunchContext * context) {
+          return NDArrayFactory::string_( shape, std::vector<const char*>(strings), dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong> &shape, const std::vector<const char *> &strings, nd4j::DataType dataType, nd4j::LaunchContext * context) {
+          std::vector<std::string> vec(strings.size());
+          int cnt = 0;
+          for (auto s:strings)
+              vec[cnt++] = std::string(s);
       
-    NDArray NDArrayFactory::string(char order, const std::vector<Nd4jLong> &shape, const std::initializer_list<const char *> &strings, nd4j::LaunchContext * context) {
+          return NDArrayFactory::string_( shape, vec, dataType, context);
-        std::vector<const char*> vec(strings);
+      }
-        return NDArrayFactory::string(order, shape, vec, context);
+      /////////////////////////////////////////////////////////////////////////
-    }
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong> &shape, const std::initializer_list<std::string> &string, nd4j::DataType dataType, nd4j::LaunchContext * context) {
-
+          return NDArrayFactory::string_( shape, std::vector<std::string>(string), dataType, context);
-    NDArray NDArrayFactory::string(char order, const std::vector<Nd4jLong> &shape, const std::vector<const char *> &strings, nd4j::LaunchContext * context) {
+      }
-        std::vector<std::string> vec(strings.size());
+      /////////////////////////////////////////////////////////////////////////
-        int cnt = 0;
+      NDArray NDArrayFactory::string( const std::vector<Nd4jLong> &shape, const std::vector<std::string> &string, nd4j::DataType dataType, nd4j::LaunchContext * context) {
-        for (auto s:strings)
+          return NDArray(shape, string, dataType, context);
-            vec[cnt++] = std::string(s);
+      }
-
+      /////////////////////////////////////////////////////////////////////////
-        return NDArrayFactory::string(order, shape, vec, context);
+      NDArray* NDArrayFactory::string_(const std::vector<Nd4jLong> &shape, const std::vector<std::string> &string, nd4j::DataType dataType, nd4j::LaunchContext * context) {
-    }
+          auto res = new NDArray();
-
+          *res = NDArray( shape, string, dataType, context);
-
+          return res;
-    NDArray NDArrayFactory::string(char order, const std::vector<Nd4jLong> &shape, const std::initializer_list<std::string> &string, nd4j::LaunchContext * context) {
+      }
-        std::vector<std::string> vec(string);
+      /////////////////////////////////////////////////////////////////////////
-        return NDArrayFactory::string(order, shape, vec, context);
+      NDArray NDArrayFactory::string(const std::vector<Nd4jLong>& shape, const std::initializer_list<const char16_t*>& strings, nd4j::DataType dataType, nd4j::LaunchContext* context) {
-    }
+          return NDArray( shape, std::vector<const char16_t*>(strings), dataType, context);
-
+      }
-    NDArray* NDArrayFactory::string_(char order, const std::vector<Nd4jLong> &shape, const std::initializer_list<const char *> &strings, nd4j::LaunchContext * context) {
+      /////////////////////////////////////////////////////////////////////////
-        std::vector<const char*> vec(strings);
+      NDArray NDArrayFactory::string( const std::vector<Nd4jLong>& shape, const std::vector<const char16_t*>& strings, nd4j::DataType dataType, nd4j::LaunchContext* context) {
-        return NDArrayFactory::string_(order, shape, vec, context);
+         return NDArray( shape, strings, dataType, context);
-    }
+      }
-
+      /////////////////////////////////////////////////////////////////////////
-    NDArray* NDArrayFactory::string_(char order, const std::vector<Nd4jLong> &shape, const std::vector<const char *> &strings, nd4j::LaunchContext * context) {
+      NDArray NDArrayFactory::string( const std::vector<Nd4jLong>& shape, const std::initializer_list<std::u16string>& string, nd4j::DataType dataType, nd4j::LaunchContext* context) {          
-        std::vector<std::string> vec(strings.size());
+          return NDArray( shape, std::vector<std::u16string>(string), dataType, context);
-        int cnt = 0;
+      }
-        for (auto s:strings)
+      /////////////////////////////////////////////////////////////////////////
-            vec[cnt++] = std::string(s);
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong>& shape, const std::initializer_list<const char16_t*>& strings, nd4j::DataType dataType, nd4j::LaunchContext* context) {
-
+          return NDArrayFactory::string_( shape, std::vector<const char16_t*>(strings), dataType, context);
-        return NDArrayFactory::string_(order, shape, vec, context);
+      }
-    }
+      /////////////////////////////////////////////////////////////////////////
-
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong>& shape, const std::vector<const char16_t*>& strings, nd4j::DataType dataType, nd4j::LaunchContext* context) {
-
+          std::vector<std::u16string> vec(strings.size());
-    NDArray* NDArrayFactory::string_(char order, const std::vector<Nd4jLong> &shape, const std::initializer_list<std::string> &string, nd4j::LaunchContext * context) {
+          int cnt = 0;
-        std::vector<std::string> vec(string);
+          for (auto s : strings)
-        return NDArrayFactory::string_(order, shape, vec, context);
+              vec[cnt++] = std::u16string(s);
-    }
-
-    NDArray NDArrayFactory::string(char order, const std::vector<Nd4jLong> &shape, const std::vector<std::string> &string, nd4j::LaunchContext * context) {
-
-          if (context == nullptr)
-            context = nd4j::LaunchContext ::defaultContext();
-
-        auto headerLength = ShapeUtils::stringBufferHeaderRequirements(string.size());
-
-        std::vector<Nd4jLong> offsets(string.size() + 1);
-        Nd4jLong dataLength = 0;
-        for (int e = 0; e < string.size(); e++) {
-            offsets[e] = dataLength;
-            dataLength += string[e].length();
-        }
-        offsets[string.size()] = dataLength;
-
-        std::shared_ptr<DataBuffer> pBuffer = std::make_shared<DataBuffer>(headerLength + dataLength, DataType::UTF8, context->getWorkspace(), true);
-
-        NDArray res(pBuffer, ShapeDescriptor(DataType::UTF8, order, shape), context);
-        res.setAttached(context->getWorkspace() != nullptr);
-
-        if (res.lengthOf() != string.size())
-            throw std::invalid_argument("Number of strings should match length of array");
-
-        memcpy(res.buffer(), offsets.data(), offsets.size() * sizeof(Nd4jLong));
-
-        auto data = static_cast<int8_t*>(res.buffer()) + headerLength;
-        int resLen = res.lengthOf();
-        for (int e = 0; e < resLen; e++) {
-            auto length = offsets[e+1] - offsets[e];
-            auto cdata = data + offsets[e];
-            memcpy(cdata, string[e].c_str(), string[e].length());
-        }
-
-        res.tickWriteHost();
-        res.syncToDevice();
-
-        return res;
-    }
-
-    NDArray* NDArrayFactory::string_(char order, const std::vector<Nd4jLong> &shape, const std::vector<std::string> &string, nd4j::LaunchContext * context) {
-        auto res = new NDArray();
-        *res = NDArrayFactory::string(order, shape, string, context);
-        return res;
-    }
       
+          return NDArrayFactory::string_( shape, vec, dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong>& shape, const std::initializer_list<std::u16string>& string, nd4j::DataType dataType, nd4j::LaunchContext* context) {
+          return NDArrayFactory::string_( shape, std::vector<std::u16string>(string), dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong>& shape, const std::vector<std::u16string>& string, nd4j::DataType dataType, nd4j::LaunchContext* context) { 
+          auto res = new NDArray();
+          *res = NDArray( shape, string, dataType, context);
+          return res;
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string( const std::vector<Nd4jLong>& shape, const std::vector<std::u16string>& string, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return NDArray( shape, string, dtype, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string( const std::vector<Nd4jLong>& shape, const std::initializer_list<const char32_t*>& strings, nd4j::DataType dataType, nd4j::LaunchContext* context) {
+          return NDArray( shape, std::vector<const char32_t*>(strings), dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string( const std::vector<Nd4jLong>& shape, const std::vector<const char32_t*>& strings, nd4j::DataType dataType, nd4j::LaunchContext* context) {
+          return NDArray( shape, strings, dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string( const std::vector<Nd4jLong>& shape, const std::initializer_list<std::u32string>& string, nd4j::DataType dataType, nd4j::LaunchContext* context) {        
+          return NDArray(shape, std::vector<std::u32string>(string), dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong>& shape, const std::initializer_list<const char32_t*>& strings, nd4j::DataType dataType, nd4j::LaunchContext* context) {
+          return NDArrayFactory::string_( shape, std::vector<const char32_t*>(strings), dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong>& shape, const std::vector<const char32_t*>& strings, nd4j::DataType dataType, nd4j::LaunchContext* context) {
+          std::vector<std::u32string> vec(strings.size());
+          int cnt = 0;
+          for (auto s : strings)
+              vec[cnt++] = std::u32string(s);      
+          return NDArrayFactory::string_( shape, vec, dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong>& shape, const std::initializer_list<std::u32string>& string, nd4j::DataType dataType, nd4j::LaunchContext* context) {
+          return NDArrayFactory::string_( shape, std::vector<std::u32string>(string), dataType, context);
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray* NDArrayFactory::string_( const std::vector<Nd4jLong>& shape, const std::vector<std::u32string>& string, nd4j::DataType dataType, nd4j::LaunchContext* context) {
+          auto res = new NDArray();
+          *res = NDArray( shape, string, dataType, context);
+          return res;
+      }
+      /////////////////////////////////////////////////////////////////////////
+      NDArray NDArrayFactory::string(const std::vector<Nd4jLong>& shape, const std::vector<std::u32string>& string, nd4j::DataType dtype, nd4j::LaunchContext* context) {
+          return NDArray( shape, string, dtype, context);
+      }
 
 }

libnd4j/blas/cpu/NativeOps.cpp

@@ -1974,7 +1974,7 @@ void deleteShapeList(Nd4jPointer shapeList) {
     delete list;
 }
 
-nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::DeclarableOp* op, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs) {
+nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::DeclarableOp* op, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs, int *dArgs, int numDArgs) {
     nd4j::graph::VariableSpace varSpace;
     Context block(2, &varSpace);
     nd4j::ShapeList inShapes;
@@ -1988,6 +1988,9 @@ nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::D
     for (int e = 0; e < numBArgs; e++)
         block.getBArguments()->push_back(bArgs[e]);
 
+    for (int e = 0; e < numDArgs; e++)
+        block.getDArguments()->push_back((nd4j::DataType) dArgs[e]);
+
     for (int e = 0; e < numInputShapes; e++) {
         auto shape_ = reinterpret_cast<Nd4jLong *>(inputShapes[e]);
 
@@ -2015,11 +2018,11 @@ nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::D
     return shapeList;
 }
 
-nd4j::ShapeList* calculateOutputShapes2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs) {
+nd4j::ShapeList* calculateOutputShapes2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs, int *dArgs, int numDArgs) {
     try {
         auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
 
-        return _calculateOutputShapes(extraPointers, op, inputBuffers, inputShapes, numInputShapes, tArgs, numTArgs, iArgs, numIArgs, bArgs, numBArgs);
+        return _calculateOutputShapes(extraPointers, op, inputBuffers, inputShapes, numInputShapes, tArgs, numTArgs, iArgs, numIArgs, bArgs, numBArgs, dArgs, numDArgs);
     } catch (std::exception &e) {
         nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
         nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
@@ -2130,7 +2133,7 @@ Nd4jStatus realExec(nd4j::ops::DeclarableOp* op, Nd4jPointer* extraPointers, Nd4
         biArgs[e] = bArgs[e];
 
     // hypothetically at this point we have everything filled
-    auto hZ = op->execute(inputs, outputs, ttArgs, iiArgs, biArgs, isInplace);
+    auto hZ = op->execute(inputs, outputs, ttArgs, iiArgs, biArgs, std::vector<nd4j::DataType>(), isInplace);
     //auto hZ = op->execute(inputs, ttArgs, iiArgs, isInplace);
 
 
@@ -2788,6 +2791,15 @@ void setGraphContextIArguments(nd4j::graph::Context* ptr, Nd4jLong *arguments, i
 void setGraphContextBArguments(nd4j::graph::Context* ptr, bool *arguments, int numberOfArguments) {
     ptr->setBArguments(arguments, numberOfArguments);
 }
+
+void setGraphContextDArguments(OpaqueContext* ptr, int *arguments, int numberOfArguments) {
+    std::vector<nd4j::DataType> dtypes(numberOfArguments);
+    for (int e = 0; e < numberOfArguments; e++)
+        dtypes[e] = (nd4j::DataType) arguments[e];
+
+    ptr->setDArguments(dtypes);
+}
+
 void deleteGraphContext(nd4j::graph::Context* ptr) {
     delete ptr;
 }

libnd4j/blas/cuda/NativeOps.cu

@@ -2684,7 +2684,7 @@ const char* getAllCustomOps() {
 }
 
 
-nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::DeclarableOp* op, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs) {
+nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::DeclarableOp* op, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs, int *dArgs, int numDArgs) {
     nd4j::graph::VariableSpace varSpace;
     Context block(2, &varSpace);
     nd4j::ShapeList inShapes;
@@ -2698,6 +2698,9 @@ nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::D
 	for (int e = 0; e < numBArgs; e++)
 		block.getBArguments()->push_back(bArgs[e]);
 
+    for (int e = 0; e < numDArgs; e++)
+        block.getDArguments()->push_back((nd4j::DataType) dArgs[e]);
+
 	for (int e = 0; e < numInputShapes; e++) {
 		auto shape_ = reinterpret_cast<Nd4jLong *>(inputShapes[e]);
 
@@ -2722,12 +2725,12 @@ nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::D
     return shapeList;
 }
 
-nd4j::ShapeList* calculateOutputShapes2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs) {
+nd4j::ShapeList* calculateOutputShapes2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs, int *dArgs, int numDArgs) {
     try {
         auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
 
         return _calculateOutputShapes(extraPointers, op, inputBuffers, inputShapes, numInputShapes, tArgs, numTArgs,
-                                      iArgs, numIArgs, bArgs, numBArgs);
+                                      iArgs, numIArgs, bArgs, numBArgs, dArgs, numDArgs);
     } catch (std::exception &e) {
         nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
         nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
@@ -2831,7 +2834,7 @@ static FORCEINLINE Nd4jStatus realExec(nd4j::ops::DeclarableOp* op, Nd4jPointer*
 
 
 	// hypothetically at this point we have everything filled
-	auto dZ = op->execute(inputs, outputs, ttArgs, iiArgs, bbArgs, isInplace);
+	auto dZ = op->execute(inputs, outputs, ttArgs, iiArgs, bbArgs, std::vector<nd4j::DataType>(), isInplace);
 	//auto dZ = op->execute(inputs, ttArgs, iiArgs, isInplace);
 
 
@@ -3596,6 +3599,14 @@ void setGraphContextBArguments(nd4j::graph::Context* ptr, bool *arguments, int n
     ptr->setBArguments(arguments, numberOfArguments);
 }
 
+void setGraphContextDArguments(OpaqueContext* ptr, int *arguments, int numberOfArguments) {
+    std::vector<nd4j::DataType> dtypes(numberOfArguments);
+    for (int e = 0; e < numberOfArguments; e++)
+        dtypes[e] = (nd4j::DataType) arguments[e];
+
+    ptr->setDArguments(dtypes);
+}
+
 void deleteGraphContext(nd4j::graph::Context* ptr) {
     delete ptr;
 }

libnd4j/buildnativeoperations.sh

@@ -55,6 +55,7 @@ TESTS="false"
 VERBOSE="false"
 VERBOSE_ARG="VERBOSE=1"
 HELPER=
+CHECK_VECTORIZATION="OFF"
 NAME=
 while [[ $# > 0 ]]
 do
@@ -114,6 +115,9 @@ case $key in
     NAME="$value"
     shift # past argument
     ;;
+    --check-vectorization)
+    CHECK_VECTORIZATION="ON"
+    ;;
     -j)
     MAKEJ="$value"
     shift # past argument
@@ -528,14 +532,27 @@ echo MINIFIER = "${MINIFIER_ARG}"
 echo TESTS = "${TESTS_ARG}"
 echo NAME = "${NAME_ARG}"
 echo OPENBLAS_PATH = "$OPENBLAS_PATH"
+echo CHECK_VECTORIZATION = "$CHECK_VECTORIZATION"
 echo HELPERS = "$HELPERS"
 mkbuilddir
 pwd
-eval $CMAKE_COMMAND  "$BLAS_ARG" "$ARCH_ARG" "$NAME_ARG" $HELPERS "$SHARED_LIBS_ARG" "$MINIFIER_ARG" "$OPERATIONS_ARG" "$BUILD_TYPE" "$PACKAGING_ARG" "$EXPERIMENTAL_ARG" "$TESTS_ARG" "$CUDA_COMPUTE" -DOPENBLAS_PATH="$OPENBLAS_PATH" -DDEV=FALSE -DCMAKE_NEED_RESPONSE=YES -DMKL_MULTI_THREADED=TRUE ../..
+eval $CMAKE_COMMAND  "$BLAS_ARG" "$ARCH_ARG" "$NAME_ARG" -DCHECK_VECTORIZATION="${CHECK_VECTORIZATION}"  $HELPERS "$SHARED_LIBS_ARG" "$MINIFIER_ARG" "$OPERATIONS_ARG" "$BUILD_TYPE" "$PACKAGING_ARG" "$EXPERIMENTAL_ARG" "$TESTS_ARG" "$CUDA_COMPUTE" -DOPENBLAS_PATH="$OPENBLAS_PATH" -DDEV=FALSE -DCMAKE_NEED_RESPONSE=YES -DMKL_MULTI_THREADED=TRUE ../..
+
 if [ "$PARALLEL" == "true" ]; then
     MAKE_ARGUMENTS="$MAKE_ARGUMENTS -j $MAKEJ"
 fi
 if [ "$VERBOSE" == "true" ]; then
     MAKE_ARGUMENTS="$MAKE_ARGUMENTS $VERBOSE_ARG"
 fi
-eval $MAKE_COMMAND $MAKE_ARGUMENTS && cd ../../..
+
+if [ "$CHECK_VECTORIZATION" == "ON" ]; then
+
+if [ "$MAKE_COMMAND" == "make" ]; then
+    MAKE_ARGUMENTS="$MAKE_ARGUMENTS --output-sync=target"
+fi
+exec 3>&1 
+eval $MAKE_COMMAND $MAKE_ARGUMENTS 2>&1 >&3 3>&-  | python3 ../../auto_vectorization/auto_vect.py   && cd ../../..
+exec 3>&- 
+else
+eval $MAKE_COMMAND $MAKE_ARGUMENTS  && cd ../../..
+fi

libnd4j/include/array/DataTypeUtils.h

@@ -95,6 +95,10 @@ namespace nd4j {
         template<typename T>
         // struct scalarTypesForNDarray { static bool const value = std::is_same<double, T>::value || std::is_same<float, T>::value || std::is_same<int, T>::value || std::is_same<bfloat16, T>::value || std::is_same<float16, T>::value || std::is_same<long long, T>::value; };
         struct scalarTypesForNDarray { static bool const value = std::is_same<double, T>::value || std::is_same<float, T>::value || std::is_same<int, T>::value || std::is_same<unsigned int, T>::value || std::is_same<long long, T>::value || std::is_same<unsigned long long, T>::value || std::is_same<long int, T>::value || std::is_same<long unsigned int, T>::value || std::is_same<int8_t, T>::value || std::is_same<uint8_t, T>::value || std::is_same<int16_t, T>::value || std::is_same<uint16_t, T>::value || std::is_same<bool, T>::value || std::is_same<bfloat16, T>::value || std::is_same<float16, T>::value; };
+
+        template<typename T>
+        struct scalarTypesForExecution { static bool const value = std::is_same<double, T>::value || std::is_same<float, T>::value || std::is_same<Nd4jLong, T>::value || std::is_same<int, T>::value || std::is_same<bool, T>::value; };
+
     };
 
 
@@ -118,7 +122,7 @@ namespace nd4j {
     }
 
     FORCEINLINE bool DataTypeUtils::isS(nd4j::DataType dataType) {
-        return dataType == nd4j::DataType::UTF8;
+        return dataType == nd4j::DataType::UTF8 || dataType == nd4j::DataType::UTF16 || dataType == nd4j::DataType::UTF32;
     }
 
     FORCEINLINE bool DataTypeUtils::isZ(nd4j::DataType dataType) {
@@ -366,6 +370,10 @@ FORCEINLINE std::string DataTypeUtils::asString(DataType dataType) {
             return std::string("UINT64");
         case UTF8:
             return std::string("UTF8");
+        case UTF16:
+            return std::string("UTF16");
+        case UTF32:
+            return std::string("UTF32");
         default:
             throw std::runtime_error("Unknown data type used");
     }
@@ -427,6 +435,8 @@ FORCEINLINE _CUDA_HD T DataTypeUtils::eps() {
             case nd4j::DataType::UINT16: return (size_t) 2;
 
             case nd4j::DataType::UTF8:
+            case nd4j::DataType::UTF16:
+            case nd4j::DataType::UTF32:
             case nd4j::DataType::INT32:
             case nd4j::DataType::UINT32:
             case nd4j::DataType::HALF2:
@@ -451,6 +461,10 @@ FORCEINLINE _CUDA_HD T DataTypeUtils::eps() {
             return nd4j::DataType::BOOL;
         } else if (std::is_same<T, std::string>::value) {
             return nd4j::DataType::UTF8;
+        } else if (std::is_same<T, std::u16string>::value) {
+            return nd4j::DataType::UTF16;
+        } else if (std::is_same<T, std::u32string>::value) {
+            return nd4j::DataType::UTF32;
         } else if (std::is_same<T, float>::value) {
             return nd4j::DataType::FLOAT32;
         } else if (std::is_same<T, float16>::value) {

libnd4j/include/array/impl/NDArrayList.cpp

@@ -158,7 +158,7 @@ namespace nd4j {
 
         iargs.push_back(_axis);
 
-        auto result = op.execute(inputs, {}, {}, {});
+        auto result = op.evaluate(inputs);
 
         auto array = new NDArray(result->at(0)->dup());
 

libnd4j/include/graph/Context.h

@@ -197,10 +197,12 @@ namespace nd4j {
             void setTArguments(double *arguments, int numberOfArguments);
             void setIArguments(Nd4jLong *arguments, int numberOfArguments);
             void setBArguments(bool *arguments, int numberOfArguments);
+            void setDArguments(nd4j::DataType *arguments, int numberOfArguments);
 
             void setTArguments(const std::vector<double> &tArgs);
             void setIArguments(const std::vector<Nd4jLong> &tArgs);
             void setBArguments(const std::vector<bool> &tArgs);
+            void setDArguments(const std::vector<nd4j::DataType> &dArgs);
 
             void setCudaContext(Nd4jPointer cudaStream, Nd4jPointer reductionPointer, Nd4jPointer allocationPointer);
 

libnd4j/include/graph/ContextPrototype.h

@@ -47,6 +47,9 @@ namespace nd4j {
             std::vector<int> _iArgs;
             std::vector<bool> _bArgs;
             std::vector<int> _axis;
+            std::vector<nd4j::DataType> _dArgs;
+
+            // TODO: remove this field
 			nd4j::DataType _dataType = nd4j::DataType::FLOAT32;
 			bool _isInplace;
 
@@ -93,6 +96,7 @@ namespace nd4j {
             std::vector<double>* getTArguments();
             std::vector<int>* getIArguments();
             std::vector<bool>* getBArguments();
+            std::vector<nd4j::DataType>* getDArguments();
             std::vector<int>* getAxis();
 
             samediff::Engine engine();
@@ -100,6 +104,7 @@ namespace nd4j {
             size_t numT();
             size_t numI();
             size_t numB();
+            size_t numD();
 
             std::pair<int, int>* input(int idx);
 

libnd4j/include/graph/Node.h

@@ -38,7 +38,9 @@ namespace nd4j {
 
         class ND4J_EXPORT Node {
         protected:
+            // TODO: this field must be removed
             nd4j::DataType _dataType;
+
             OpType _opType;
             ContextPrototype* _protoContext = nullptr;
             Nd4jLong _opNum;
@@ -61,6 +63,7 @@ namespace nd4j {
 
 
             // optional scalar. used in scalar ops and in summary stats
+            // TODO: this field must be removed
             NDArray _scalar;
 
             bool _hasExternalOutputs;
@@ -87,15 +90,15 @@ namespace nd4j {
             int _scope_id = 0;
             std::string _scope_name;
 
+            // TODO: these 3 fields should be removed
             int _rewindNode = -1;
             std::pair<int, int> _rewindLayer = {-1, -1};
-
             Nd4jLong _frameId = -1;
 
         public:
-            Node(nd4j::ops::DeclarableOp *customOp, int id = 0, std::initializer_list<int> input = {}, std::initializer_list<int> output = {},  std::initializer_list<int> dimensions = {}, float scalar = 0.0f, std::initializer_list<double> tArgs = {}, std::initializer_list<int> iArgs = {});
+            explicit Node(nd4j::ops::DeclarableOp *customOp, int id = 0, std::initializer_list<int> input = {}, std::initializer_list<int> output = {},  std::initializer_list<int> dimensions = {}, float scalar = 0.0f, std::initializer_list<double> tArgs = {}, std::initializer_list<int> iArgs = {});
-            Node(OpType opType = OpType_TRANSFORM_SAME, int opNum = 0, int id = 0, std::initializer_list<int> input = {}, std::initializer_list<int> output = {},  std::initializer_list<int> dimensions = {}, float scalar = 0.0f, std::initializer_list<double> tArgs = {}, std::initializer_list<int> iArgs = {});
+            explicit Node(OpType opType = OpType_TRANSFORM_SAME, int opNum = 0, int id = 0, std::initializer_list<int> input = {}, std::initializer_list<int> output = {},  std::initializer_list<int> dimensions = {}, float scalar = 0.0f, std::initializer_list<double> tArgs = {}, std::initializer_list<int> iArgs = {});
-            Node(const nd4j::graph::FlatNode *node);
+            explicit Node(const nd4j::graph::FlatNode *node);
             ~Node();
 
             bool equals(Node *other);

libnd4j/include/graph/generated/array_generated.h

@@ -60,11 +60,13 @@ enum DType {
   DType_QINT16 = 16,
   DType_BFLOAT16 = 17,
   DType_UTF8 = 50,
+  DType_UTF16 = 51,
+  DType_UTF32 = 52,
   DType_MIN = DType_INHERIT,
-  DType_MAX = DType_UTF8
+  DType_MAX = DType_UTF32
 };
 
-inline const DType (&EnumValuesDType())[19] {
+inline const DType (&EnumValuesDType())[21] {
   static const DType values[] = {
     DType_INHERIT,
     DType_BOOL,
@@ -84,7 +86,9 @@ inline const DType (&EnumValuesDType())[19] {
     DType_QINT8,
     DType_QINT16,
     DType_BFLOAT16,
-    DType_UTF8
+    DType_UTF8,
+    DType_UTF16,
+    DType_UTF32
   };
   return values;
 }
@@ -142,6 +146,8 @@ inline const char * const *EnumNamesDType() {
     "",
     "",
     "UTF8",
+    "UTF16",
+    "UTF32",
     nullptr
   };
   return names;

libnd4j/include/graph/generated/array_generated.js

@@ -42,7 +42,9 @@ nd4j.graph.DType = {
   QINT8: 15,
   QINT16: 16,
   BFLOAT16: 17,
-  UTF8: 50
+  UTF8: 50,
+  UTF16: 51,
+  UTF32: 52
 };
 
 /**

libnd4j/include/graph/generated/nd4j/graph/DType.cs

@@ -26,6 +26,8 @@ public enum DType : sbyte
  QINT16 = 16,
  BFLOAT16 = 17,
  UTF8 = 50,
+ UTF16 = 51,
+ UTF32 = 52,
 };
 
 

libnd4j/include/graph/generated/nd4j/graph/DType.java

@@ -23,8 +23,10 @@ public final class DType {
   public static final byte QINT16 = 16;
   public static final byte BFLOAT16 = 17;
   public static final byte UTF8 = 50;
+  public static final byte UTF16 = 51;
+  public static final byte UTF32 = 52;
 
-  public static final String[] names = { "INHERIT", "BOOL", "FLOAT8", "HALF", "HALF2", "FLOAT", "DOUBLE", "INT8", "INT16", "INT32", "INT64", "UINT8", "UINT16", "UINT32", "UINT64", "QINT8", "QINT16", "BFLOAT16", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "UTF8", };
+  public static final String[] names = { "INHERIT", "BOOL", "FLOAT8", "HALF", "HALF2", "FLOAT", "DOUBLE", "INT8", "INT16", "INT32", "INT64", "UINT8", "UINT16", "UINT32", "UINT64", "QINT8", "QINT16", "BFLOAT16", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "UTF8", "UTF16", "UTF32", };
 
   public static String name(int e) { return names[e]; }
 }

libnd4j/include/graph/generated/nd4j/graph/DType.py

@@ -22,4 +22,6 @@ class DType(object):
     QINT16 = 16
     BFLOAT16 = 17
     UTF8 = 50
+    UTF16 = 51
+    UTF32 = 52
 

libnd4j/include/graph/generated/nd4j/graph/FlatNode.cs

@@ -112,6 +112,14 @@ public struct FlatNode : IFlatbufferObject
   public int VarControlDepsLength { get { int o = __p.__offset(44); return o != 0 ? __p.__vector_len(o) : 0; } }
   public string ControlDepFor(int j) { int o = __p.__offset(46); return o != 0 ? __p.__string(__p.__vector(o) + j * 4) : null; }
   public int ControlDepForLength { get { int o = __p.__offset(46); return o != 0 ? __p.__vector_len(o) : 0; } }
+  public DType ExtraTypes(int j) { int o = __p.__offset(48); return o != 0 ? (DType)__p.bb.GetSbyte(__p.__vector(o) + j * 1) : (DType)0; }
+  public int ExtraTypesLength { get { int o = __p.__offset(48); return o != 0 ? __p.__vector_len(o) : 0; } }
+#if ENABLE_SPAN_T
+  public Span<byte> GetExtraTypesBytes() { return __p.__vector_as_span(48); }
+#else
+  public ArraySegment<byte>? GetExtraTypesBytes() { return __p.__vector_as_arraysegment(48); }
+#endif
+  public DType[] GetExtraTypesArray() { return __p.__vector_as_array<DType>(48); }
 
   public static Offset<FlatNode> CreateFlatNode(FlatBufferBuilder builder,
       int id = 0,
@@ -135,9 +143,11 @@ public struct FlatNode : IFlatbufferObject
       Offset<FlatArray> scalarOffset = default(Offset<FlatArray>),
       VectorOffset controlDepsOffset = default(VectorOffset),
       VectorOffset varControlDepsOffset = default(VectorOffset),
-      VectorOffset controlDepForOffset = default(VectorOffset)) {
+      VectorOffset controlDepForOffset = default(VectorOffset),
-    builder.StartObject(22);
+      VectorOffset extraTypesOffset = default(VectorOffset)) {
+    builder.StartObject(23);
     FlatNode.AddOpNum(builder, opNum);
+    FlatNode.AddExtraTypes(builder, extraTypesOffset);
     FlatNode.AddControlDepFor(builder, controlDepForOffset);
     FlatNode.AddVarControlDeps(builder, varControlDepsOffset);
     FlatNode.AddControlDeps(builder, controlDepsOffset);
@@ -162,7 +172,7 @@ public struct FlatNode : IFlatbufferObject
     return FlatNode.EndFlatNode(builder);
   }
 
-  public static void StartFlatNode(FlatBufferBuilder builder) { builder.StartObject(22); }
+  public static void StartFlatNode(FlatBufferBuilder builder) { builder.StartObject(23); }
   public static void AddId(FlatBufferBuilder builder, int id) { builder.AddInt(0, id, 0); }
   public static void AddName(FlatBufferBuilder builder, StringOffset nameOffset) { builder.AddOffset(1, nameOffset.Value, 0); }
   public static void AddOpType(FlatBufferBuilder builder, OpType opType) { builder.AddSbyte(2, (sbyte)opType, 0); }
@@ -224,6 +234,10 @@ public struct FlatNode : IFlatbufferObject
   public static VectorOffset CreateControlDepForVector(FlatBufferBuilder builder, StringOffset[] data) { builder.StartVector(4, data.Length, 4); for (int i = data.Length - 1; i >= 0; i--) builder.AddOffset(data[i].Value); return builder.EndVector(); }
   public static VectorOffset CreateControlDepForVectorBlock(FlatBufferBuilder builder, StringOffset[] data) { builder.StartVector(4, data.Length, 4); builder.Add(data); return builder.EndVector(); }
   public static void StartControlDepForVector(FlatBufferBuilder builder, int numElems) { builder.StartVector(4, numElems, 4); }
+  public static void AddExtraTypes(FlatBufferBuilder builder, VectorOffset extraTypesOffset) { builder.AddOffset(22, extraTypesOffset.Value, 0); }
+  public static VectorOffset CreateExtraTypesVector(FlatBufferBuilder builder, DType[] data) { builder.StartVector(1, data.Length, 1); for (int i = data.Length - 1; i >= 0; i--) builder.AddSbyte((sbyte)data[i]); return builder.EndVector(); }
+  public static VectorOffset CreateExtraTypesVectorBlock(FlatBufferBuilder builder, DType[] data) { builder.StartVector(1, data.Length, 1); builder.Add(data); return builder.EndVector(); }
+  public static void StartExtraTypesVector(FlatBufferBuilder builder, int numElems) { builder.StartVector(1, numElems, 1); }
   public static Offset<FlatNode> EndFlatNode(FlatBufferBuilder builder) {
     int o = builder.EndObject();
     return new Offset<FlatNode>(o);

libnd4j/include/graph/generated/nd4j/graph/FlatNode.java

@@ -72,6 +72,10 @@ public final class FlatNode extends Table {
   public int varControlDepsLength() { int o = __offset(44); return o != 0 ? __vector_len(o) : 0; }
   public String controlDepFor(int j) { int o = __offset(46); return o != 0 ? __string(__vector(o) + j * 4) : null; }
   public int controlDepForLength() { int o = __offset(46); return o != 0 ? __vector_len(o) : 0; }
+  public byte extraTypes(int j) { int o = __offset(48); return o != 0 ? bb.get(__vector(o) + j * 1) : 0; }
+  public int extraTypesLength() { int o = __offset(48); return o != 0 ? __vector_len(o) : 0; }
+  public ByteBuffer extraTypesAsByteBuffer() { return __vector_as_bytebuffer(48, 1); }
+  public ByteBuffer extraTypesInByteBuffer(ByteBuffer _bb) { return __vector_in_bytebuffer(_bb, 48, 1); }
 
   public static int createFlatNode(FlatBufferBuilder builder,
       int id,
@@ -95,9 +99,11 @@ public final class FlatNode extends Table {
       int scalarOffset,
       int controlDepsOffset,
       int varControlDepsOffset,
-      int controlDepForOffset) {
+      int controlDepForOffset,
-    builder.startObject(22);
+      int extraTypesOffset) {
+    builder.startObject(23);
     FlatNode.addOpNum(builder, opNum);
+    FlatNode.addExtraTypes(builder, extraTypesOffset);
     FlatNode.addControlDepFor(builder, controlDepForOffset);
     FlatNode.addVarControlDeps(builder, varControlDepsOffset);
     FlatNode.addControlDeps(builder, controlDepsOffset);
@@ -122,7 +128,7 @@ public final class FlatNode extends Table {
     return FlatNode.endFlatNode(builder);
   }
 
-  public static void startFlatNode(FlatBufferBuilder builder) { builder.startObject(22); }
+  public static void startFlatNode(FlatBufferBuilder builder) { builder.startObject(23); }
   public static void addId(FlatBufferBuilder builder, int id) { builder.addInt(0, id, 0); }
   public static void addName(FlatBufferBuilder builder, int nameOffset) { builder.addOffset(1, nameOffset, 0); }
   public static void addOpType(FlatBufferBuilder builder, byte opType) { builder.addByte(2, opType, 0); }
@@ -171,6 +177,9 @@ public final class FlatNode extends Table {
   public static void addControlDepFor(FlatBufferBuilder builder, int controlDepForOffset) { builder.addOffset(21, controlDepForOffset, 0); }
   public static int createControlDepForVector(FlatBufferBuilder builder, int[] data) { builder.startVector(4, data.length, 4); for (int i = data.length - 1; i >= 0; i--) builder.addOffset(data[i]); return builder.endVector(); }
   public static void startControlDepForVector(FlatBufferBuilder builder, int numElems) { builder.startVector(4, numElems, 4); }
+  public static void addExtraTypes(FlatBufferBuilder builder, int extraTypesOffset) { builder.addOffset(22, extraTypesOffset, 0); }
+  public static int createExtraTypesVector(FlatBufferBuilder builder, byte[] data) { builder.startVector(1, data.length, 1); for (int i = data.length - 1; i >= 0; i--) builder.addByte(data[i]); return builder.endVector(); }
+  public static void startExtraTypesVector(FlatBufferBuilder builder, int numElems) { builder.startVector(1, numElems, 1); }
   public static int endFlatNode(FlatBufferBuilder builder) {
     int o = builder.endObject();
     return o;

libnd4j/include/graph/generated/nd4j/graph/FlatNode.py

@@ -339,7 +339,29 @@ class FlatNode(object):
             return self._tab.VectorLen(o)
         return 0
 
-def FlatNodeStart(builder): builder.StartObject(22)
+    # FlatNode
+    def ExtraTypes(self, j):
+        o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(48))
+        if o != 0:
+            a = self._tab.Vector(o)
+            return self._tab.Get(flatbuffers.number_types.Int8Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 1))
+        return 0
+
+    # FlatNode
+    def ExtraTypesAsNumpy(self):
+        o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(48))
+        if o != 0:
+            return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int8Flags, o)
+        return 0
+
+    # FlatNode
+    def ExtraTypesLength(self):
+        o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(48))
+        if o != 0:
+            return self._tab.VectorLen(o)
+        return 0
+
+def FlatNodeStart(builder): builder.StartObject(23)
 def FlatNodeAddId(builder, id): builder.PrependInt32Slot(0, id, 0)
 def FlatNodeAddName(builder, name): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(name), 0)
 def FlatNodeAddOpType(builder, opType): builder.PrependInt8Slot(2, opType, 0)
@@ -375,4 +397,6 @@ def FlatNodeAddVarControlDeps(builder, varControlDeps): builder.PrependUOffsetTR
 def FlatNodeStartVarControlDepsVector(builder, numElems): return builder.StartVector(4, numElems, 4)
 def FlatNodeAddControlDepFor(builder, controlDepFor): builder.PrependUOffsetTRelativeSlot(21, flatbuffers.number_types.UOffsetTFlags.py_type(controlDepFor), 0)
 def FlatNodeStartControlDepForVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def FlatNodeAddExtraTypes(builder, extraTypes): builder.PrependUOffsetTRelativeSlot(22, flatbuffers.number_types.UOffsetTFlags.py_type(extraTypes), 0)
+def FlatNodeStartExtraTypesVector(builder, numElems): return builder.StartVector(1, numElems, 1)
 def FlatNodeEnd(builder): return builder.EndObject()

libnd4j/include/graph/generated/nd4j/graph/UIVariable.cs

@@ -26,7 +26,7 @@ public struct UIVariable : IFlatbufferObject
 #endif
   public byte[] GetNameArray() { return __p.__vector_as_array<byte>(6); }
   public VarType Type { get { int o = __p.__offset(8); return o != 0 ? (VarType)__p.bb.GetSbyte(o + __p.bb_pos) : VarType.VARIABLE; } }
-  public DataType Datatype { get { int o = __p.__offset(10); return o != 0 ? (DataType)__p.bb.GetSbyte(o + __p.bb_pos) : DataType.INHERIT; } }
+  public DType Datatype { get { int o = __p.__offset(10); return o != 0 ? (DType)__p.bb.GetSbyte(o + __p.bb_pos) : DType.INHERIT; } }
   public long Shape(int j) { int o = __p.__offset(12); return o != 0 ? __p.bb.GetLong(__p.__vector(o) + j * 8) : (long)0; }
   public int ShapeLength { get { int o = __p.__offset(12); return o != 0 ? __p.__vector_len(o) : 0; } }
 #if ENABLE_SPAN_T
@@ -70,7 +70,7 @@ public struct UIVariable : IFlatbufferObject
       Offset<IntPair> idOffset = default(Offset<IntPair>),
       StringOffset nameOffset = default(StringOffset),
       VarType type = VarType.VARIABLE,
-      DataType datatype = DataType.INHERIT,
+      DType datatype = DType.INHERIT,
       VectorOffset shapeOffset = default(VectorOffset),
       VectorOffset controlDepsOffset = default(VectorOffset),
       StringOffset outputOfOpOffset = default(StringOffset),
@@ -101,7 +101,7 @@ public struct UIVariable : IFlatbufferObject
   public static void AddId(FlatBufferBuilder builder, Offset<IntPair> idOffset) { builder.AddOffset(0, idOffset.Value, 0); }
   public static void AddName(FlatBufferBuilder builder, StringOffset nameOffset) { builder.AddOffset(1, nameOffset.Value, 0); }
   public static void AddType(FlatBufferBuilder builder, VarType type) { builder.AddSbyte(2, (sbyte)type, 0); }
-  public static void AddDatatype(FlatBufferBuilder builder, DataType datatype) { builder.AddSbyte(3, (sbyte)datatype, 0); }
+  public static void AddDatatype(FlatBufferBuilder builder, DType datatype) { builder.AddSbyte(3, (sbyte)datatype, 0); }
   public static void AddShape(FlatBufferBuilder builder, VectorOffset shapeOffset) { builder.AddOffset(4, shapeOffset.Value, 0); }
   public static VectorOffset CreateShapeVector(FlatBufferBuilder builder, long[] data) { builder.StartVector(8, data.Length, 8); for (int i = data.Length - 1; i >= 0; i--) builder.AddLong(data[i]); return builder.EndVector(); }
   public static VectorOffset CreateShapeVectorBlock(FlatBufferBuilder builder, long[] data) { builder.StartVector(8, data.Length, 8); builder.Add(data); return builder.EndVector(); }

libnd4j/include/graph/generated/node_generated.h

@@ -38,7 +38,8 @@ struct FlatNode FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
     VT_SCALAR = 40,
     VT_CONTROLDEPS = 42,
     VT_VARCONTROLDEPS = 44,
-    VT_CONTROLDEPFOR = 46
+    VT_CONTROLDEPFOR = 46,
+    VT_EXTRATYPES = 48
   };
   int32_t id() const {
     return GetField<int32_t>(VT_ID, 0);
@@ -106,6 +107,9 @@ struct FlatNode FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
   const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *controlDepFor() const {
     return GetPointer<const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *>(VT_CONTROLDEPFOR);
   }
+  const flatbuffers::Vector<int8_t> *extraTypes() const {
+    return GetPointer<const flatbuffers::Vector<int8_t> *>(VT_EXTRATYPES);
+  }
   bool Verify(flatbuffers::Verifier &verifier) const {
     return VerifyTableStart(verifier) &&
            VerifyField<int32_t>(verifier, VT_ID) &&
@@ -153,6 +157,8 @@ struct FlatNode FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
            VerifyOffset(verifier, VT_CONTROLDEPFOR) &&
            verifier.VerifyVector(controlDepFor()) &&
            verifier.VerifyVectorOfStrings(controlDepFor()) &&
+           VerifyOffset(verifier, VT_EXTRATYPES) &&
+           verifier.VerifyVector(extraTypes()) &&
            verifier.EndTable();
   }
 };
@@ -226,6 +232,9 @@ struct FlatNodeBuilder {
   void add_controlDepFor(flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> controlDepFor) {
     fbb_.AddOffset(FlatNode::VT_CONTROLDEPFOR, controlDepFor);
   }
+  void add_extraTypes(flatbuffers::Offset<flatbuffers::Vector<int8_t>> extraTypes) {
+    fbb_.AddOffset(FlatNode::VT_EXTRATYPES, extraTypes);
+  }
   explicit FlatNodeBuilder(flatbuffers::FlatBufferBuilder &_fbb)
         : fbb_(_fbb) {
     start_ = fbb_.StartTable();
@@ -261,9 +270,11 @@ inline flatbuffers::Offset<FlatNode> CreateFlatNode(
     flatbuffers::Offset<FlatArray> scalar = 0,
     flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> controlDeps = 0,
     flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> varControlDeps = 0,
-    flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> controlDepFor = 0) {
+    flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> controlDepFor = 0,
+    flatbuffers::Offset<flatbuffers::Vector<int8_t>> extraTypes = 0) {
   FlatNodeBuilder builder_(_fbb);
   builder_.add_opNum(opNum);
+  builder_.add_extraTypes(extraTypes);
   builder_.add_controlDepFor(controlDepFor);
   builder_.add_varControlDeps(varControlDeps);
   builder_.add_controlDeps(controlDeps);
@@ -311,7 +322,8 @@ inline flatbuffers::Offset<FlatNode> CreateFlatNodeDirect(
     flatbuffers::Offset<FlatArray> scalar = 0,
     const std::vector<flatbuffers::Offset<flatbuffers::String>> *controlDeps = nullptr,
     const std::vector<flatbuffers::Offset<flatbuffers::String>> *varControlDeps = nullptr,
-    const std::vector<flatbuffers::Offset<flatbuffers::String>> *controlDepFor = nullptr) {
+    const std::vector<flatbuffers::Offset<flatbuffers::String>> *controlDepFor = nullptr,
+    const std::vector<int8_t> *extraTypes = nullptr) {
   return nd4j::graph::CreateFlatNode(
       _fbb,
       id,
@@ -335,7 +347,8 @@ inline flatbuffers::Offset<FlatNode> CreateFlatNodeDirect(
       scalar,
       controlDeps ? _fbb.CreateVector<flatbuffers::Offset<flatbuffers::String>>(*controlDeps) : 0,
       varControlDeps ? _fbb.CreateVector<flatbuffers::Offset<flatbuffers::String>>(*varControlDeps) : 0,
-      controlDepFor ? _fbb.CreateVector<flatbuffers::Offset<flatbuffers::String>>(*controlDepFor) : 0);
+      controlDepFor ? _fbb.CreateVector<flatbuffers::Offset<flatbuffers::String>>(*controlDepFor) : 0,
+      extraTypes ? _fbb.CreateVector<int8_t>(*extraTypes) : 0);
 }
 
 inline const nd4j::graph::FlatNode *GetFlatNode(const void *buf) {

libnd4j/include/graph/generated/node_generated.js

@@ -398,11 +398,36 @@ nd4j.graph.FlatNode.prototype.controlDepForLength = function() {
   return offset ? this.bb.__vector_len(this.bb_pos + offset) : 0;
 };
 
+/**
+ * @param {number} index
+ * @returns {nd4j.graph.DType}
+ */
+nd4j.graph.FlatNode.prototype.extraTypes = function(index) {
+  var offset = this.bb.__offset(this.bb_pos, 48);
+  return offset ? /** @type {nd4j.graph.DType} */ (this.bb.readInt8(this.bb.__vector(this.bb_pos + offset) + index)) : /** @type {nd4j.graph.DType} */ (0);
+};
+
+/**
+ * @returns {number}
+ */
+nd4j.graph.FlatNode.prototype.extraTypesLength = function() {
+  var offset = this.bb.__offset(this.bb_pos, 48);
+  return offset ? this.bb.__vector_len(this.bb_pos + offset) : 0;
+};
+
+/**
+ * @returns {Int8Array}
+ */
+nd4j.graph.FlatNode.prototype.extraTypesArray = function() {
+  var offset = this.bb.__offset(this.bb_pos, 48);
+  return offset ? new Int8Array(this.bb.bytes().buffer, this.bb.bytes().byteOffset + this.bb.__vector(this.bb_pos + offset), this.bb.__vector_len(this.bb_pos + offset)) : null;
+};
+
 /**
  * @param {flatbuffers.Builder} builder
  */
 nd4j.graph.FlatNode.startFlatNode = function(builder) {
-  builder.startObject(22);
+  builder.startObject(23);
 };
 
 /**
@@ -854,6 +879,35 @@ nd4j.graph.FlatNode.startControlDepForVector = function(builder, numElems) {
   builder.startVector(4, numElems, 4);
 };
 
+/**
+ * @param {flatbuffers.Builder} builder
+ * @param {flatbuffers.Offset} extraTypesOffset
+ */
+nd4j.graph.FlatNode.addExtraTypes = function(builder, extraTypesOffset) {
+  builder.addFieldOffset(22, extraTypesOffset, 0);
+};
+
+/**
+ * @param {flatbuffers.Builder} builder
+ * @param {Array.<nd4j.graph.DType>} data
+ * @returns {flatbuffers.Offset}
+ */
+nd4j.graph.FlatNode.createExtraTypesVector = function(builder, data) {
+  builder.startVector(1, data.length, 1);
+  for (var i = data.length - 1; i >= 0; i--) {
+    builder.addInt8(data[i]);
+  }
+  return builder.endVector();
+};
+
+/**
+ * @param {flatbuffers.Builder} builder
+ * @param {number} numElems
+ */
+nd4j.graph.FlatNode.startExtraTypesVector = function(builder, numElems) {
+  builder.startVector(1, numElems, 1);
+};
+
 /**
  * @param {flatbuffers.Builder} builder
  * @returns {flatbuffers.Offset}

libnd4j/include/graph/generated/uigraphstatic_generated.h

@@ -266,8 +266,8 @@ struct UIVariable FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
   VarType type() const {
     return static_cast<VarType>(GetField<int8_t>(VT_TYPE, 0));
   }
-  DataType datatype() const {
+  DType datatype() const {
-    return static_cast<DataType>(GetField<int8_t>(VT_DATATYPE, 0));
+    return static_cast<DType>(GetField<int8_t>(VT_DATATYPE, 0));
   }
   const flatbuffers::Vector<int64_t> *shape() const {
     return GetPointer<const flatbuffers::Vector<int64_t> *>(VT_SHAPE);
@@ -342,7 +342,7 @@ struct UIVariableBuilder {
   void add_type(VarType type) {
     fbb_.AddElement<int8_t>(UIVariable::VT_TYPE, static_cast<int8_t>(type), 0);
   }
-  void add_datatype(DataType datatype) {
+  void add_datatype(DType datatype) {
     fbb_.AddElement<int8_t>(UIVariable::VT_DATATYPE, static_cast<int8_t>(datatype), 0);
   }
   void add_shape(flatbuffers::Offset<flatbuffers::Vector<int64_t>> shape) {
@@ -389,7 +389,7 @@ inline flatbuffers::Offset<UIVariable> CreateUIVariable(
     flatbuffers::Offset<IntPair> id = 0,
     flatbuffers::Offset<flatbuffers::String> name = 0,
     VarType type = VarType_VARIABLE,
-    DataType datatype = DataType_INHERIT,
+    DType datatype = DType_INHERIT,
     flatbuffers::Offset<flatbuffers::Vector<int64_t>> shape = 0,
     flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> controlDeps = 0,
     flatbuffers::Offset<flatbuffers::String> outputOfOp = 0,
@@ -421,7 +421,7 @@ inline flatbuffers::Offset<UIVariable> CreateUIVariableDirect(
     flatbuffers::Offset<IntPair> id = 0,
     const char *name = nullptr,
     VarType type = VarType_VARIABLE,
-    DataType datatype = DataType_INHERIT,
+    DType datatype = DType_INHERIT,
     const std::vector<int64_t> *shape = nullptr,
     const std::vector<flatbuffers::Offset<flatbuffers::String>> *controlDeps = nullptr,
     const char *outputOfOp = nullptr,

libnd4j/include/graph/generated/uigraphstatic_generated.js

@@ -503,11 +503,11 @@ nd4j.graph.UIVariable.prototype.type = function() {
 };
 
 /**
- * @returns {nd4j.graph.DataType}
+ * @returns {nd4j.graph.DType}
  */
 nd4j.graph.UIVariable.prototype.datatype = function() {
   var offset = this.bb.__offset(this.bb_pos, 10);
-  return offset ? /** @type {nd4j.graph.DataType} */ (this.bb.readInt8(this.bb_pos + offset)) : nd4j.graph.DataType.INHERIT;
+  return offset ? /** @type {nd4j.graph.DType} */ (this.bb.readInt8(this.bb_pos + offset)) : nd4j.graph.DType.INHERIT;
 };
 
 /**
@@ -668,10 +668,10 @@ nd4j.graph.UIVariable.addType = function(builder, type) {
 
 /**
  * @param {flatbuffers.Builder} builder
- * @param {nd4j.graph.DataType} datatype
+ * @param {nd4j.graph.DType} datatype
  */
 nd4j.graph.UIVariable.addDatatype = function(builder, datatype) {
-  builder.addFieldInt8(3, datatype, nd4j.graph.DataType.INHERIT);
+  builder.addFieldInt8(3, datatype, nd4j.graph.DType.INHERIT);
 };
 
 /**

libnd4j/include/graph/impl/Context.cpp

@@ -551,6 +551,18 @@ namespace nd4j {
         bool Context::isInference() {
             return _execMode == samediff::ExecutionMode::MODE_INFERENCE;
         }
+
+        void Context::setDArguments(nd4j::DataType *arguments, int numberOfArguments) {
+            _dArgs.clear();
+            for (int e = 0; e < numberOfArguments; e++)
+                _dArgs.emplace_back(arguments[e]);
+        }
+
+        void Context::setDArguments(const std::vector<nd4j::DataType> &dArgs) {
+            _dArgs.clear();
+            for (auto d:dArgs)
+                _dArgs.emplace_back(d);
+        }
     }
 }
 

libnd4j/include/graph/impl/ContextPrototype.cpp

@@ -173,5 +173,13 @@ namespace nd4j {
 
             return clone;
         }
+
+        std::vector<nd4j::DataType> *ContextPrototype::getDArguments() {
+            return &_dArgs;
+        }
+
+        size_t ContextPrototype::numD() {
+            return _dArgs.size();
+        }
     }
 }

libnd4j/include/graph/impl/FlatUtils.cpp

@@ -49,11 +49,10 @@ namespace nd4j {
                 delete[] newShape;
                 return NDArrayFactory::empty_(dtype, nullptr);
             }
-
+            // TODO fix UTF16 and UTF32
             if (dtype == UTF8) {
                 bool isBe = BitwiseUtils::isBE();
                 bool canKeep = (isBe && flatArray->byteOrder() == nd4j::graph::ByteOrder_BE) || (!isBe && flatArray->byteOrder() == nd4j::graph::ByteOrder_LE);
-                auto order = shape::order(newShape);
                 
                 std::vector<std::string> substrings(length);
                 std::vector<Nd4jLong> shapeVector(rank);
@@ -88,8 +87,8 @@ namespace nd4j {
 
                 delete[] offsets;
                 delete[] newShape;
-
+                // string order always 'c'
-                return NDArrayFactory::string_(order, shapeVector, substrings);
+                return NDArrayFactory::string_(shapeVector, substrings);
             }
 
 

libnd4j/include/graph/impl/Node.cpp

@@ -587,6 +587,12 @@ namespace nd4j {
                                 block->getIArguments()->emplace_back(node->extraInteger()->Get(e));
                             }
 
+                        if (node->extraTypes() != nullptr && node->extraTypes()->size() > 0) {
+                            for (int e = 0; e < (int) node->extraTypes()->size(); e++) {
+                                block->getDArguments()->emplace_back((nd4j::DataType) node->extraTypes()->Get(e));
+                            }
+                        }
+
                         this->setContextPrototype(block);
                         this->setCustomOp(Node::buildOpByType(_opType, (int) node->input()->size(), (int) block->getIArguments()->size(), (int) block->getTArguments()->size(), (int) _opNum, &_scalar));
                         block->setOpDescriptor(this->getCustomOp()->getOpDescriptor());
@@ -618,6 +624,12 @@ namespace nd4j {
                                 block->getIArguments()->emplace_back(node->extraInteger()->Get(e));
                             }
 
+                        if (node->extraTypes() != nullptr && node->extraTypes()->size() > 0) {
+                            for (int e = 0; e < (int) node->extraTypes()->size(); e++) {
+                                block->getDArguments()->emplace_back((nd4j::DataType) node->extraTypes()->Get(e));
+                            }
+                        }
+
                         this->setContextPrototype(block);
 
                         this->setCustomOp(Node::buildOpByType(_opType, (int) node->inputPaired()->size(), (int) block->getIArguments()->size(), (int) block->getTArguments()->size(), (int) _opNum, &_scalar));
@@ -652,6 +664,12 @@ namespace nd4j {
                                 block->getBArguments()->push_back(node->extraBools()->Get(e));
                             }
 
+                        if (node->extraTypes() != nullptr && node->extraTypes()->size() > 0) {
+                            for (int e = 0; e < (int) node->extraTypes()->size(); e++) {
+                                block->getDArguments()->emplace_back((nd4j::DataType) node->extraTypes()->Get(e));
+                            }
+                        }
+
                         for (auto v: _dimensions)
                             block->getAxis()->emplace_back(v);
 

libnd4j/include/graph/scheme/node.fbs

@@ -58,6 +58,8 @@ table FlatNode {
 	varControlDeps:[string];
 	controlDepFor:[string];
 
+	// DArgs
+	extraTypes:[DType];
 }
 
 root_type FlatNode;

libnd4j/include/helpers/ShapeUtils.h

@@ -171,7 +171,10 @@ namespace nd4j {
          * @param numStrings
          * @return
          */
-        static Nd4jLong stringBufferHeaderRequirements(Nd4jLong numStrings);
+        static FORCEINLINE Nd4jLong stringBufferHeaderRequirements(Nd4jLong numStrings) {
+            // we store +1 offset
+            return (numStrings + 1) * sizeof(Nd4jLong);
+        }
 
         /*
         * check whether arr1/arr2 is sub-array of arr2/arr1,

libnd4j/include/helpers/StringUtils.h

@@ -1,5 +1,6 @@
 /*******************************************************************************
  * Copyright (c) 2015-2018 Skymind, Inc.
+ * Copyright (c) 2019-2020 Konduit K.K.
  *
  * This program and the accompanying materials are made available under the
  * terms of the Apache License, Version 2.0 which is available at
@@ -16,6 +17,7 @@
 
 //
 // Created by raver119 on 20/04/18.
+// @author Oleg Semeniv <oleg.semeniv@gmail.com>
 //
 
 #ifndef LIBND4J_STRINGUTILS_H
@@ -27,6 +29,7 @@
 #include <sstream>
 #include <vector>
 #include <NDArray.h>
+#include <unicode.h>
 
 namespace nd4j {
     class ND4J_EXPORT StringUtils {
@@ -85,6 +88,55 @@ namespace nd4j {
          * @return
          */
         static std::vector<std::string> split(const std::string &haystack, const std::string &delimiter);
+
+
+        /**
+         * This method convert u8 string to u16
+         * @param const reference to input string
+         * @param reference to output u16string
+         * @return boolean status
+         */
+        static bool u8StringToU16String(const std::string& u8, std::u16string& u16);
+
+        /**
+         * This method convert u8 string to u32
+         * @param const reference to input string
+         * @param reference to output u32string
+         * @return boolean status
+         */
+        static bool u8StringToU32String(const std::string& u8, std::u32string& u32);
+
+        /**
+         * This method convert u16 string to u32
+         * @param const reference to input u16string
+         * @param reference to output u32string
+         * @return boolean status
+         */
+        static bool u16StringToU32String(const std::u16string& u16, std::u32string& u32);
+
+        /**
+         * This method convert u16 string to u8 string
+         * @param const reference to input u16string
+         * @param reference to output string
+         * @return boolean status
+         */
+        static bool u16StringToU8String(const std::u16string& u16, std::string& u8);
+
+        /**
+         * This method convert u32 string to u16 string
+         * @param const reference to input u32string
+         * @param reference to output u16string
+         * @return boolean status
+         */
+        static bool u32StringToU16String(const std::u32string& u32, std::u16string& u16);
+
+        /**
+        * This method convert u32 string to u8 string
+        * @param const reference to input u32string
+        * @param reference to output string
+        * @return boolean status
+        */
+        static bool u32StringToU8String(const std::u32string& u32, std::string& u8);
     };
 }
 

libnd4j/include/helpers/impl/AttentionHelper.cpp

@@ -40,7 +40,7 @@ namespace nd4j {
 
         NDArray projected('c', {numHeads * projectionMatrix->sizeAt(1), (miniBatchSize * seqLength)}, input->dataType(), context);  //[nHeads*hS, batch*timeSteps]
         nd4j::ops::matmul mmul;
-        mmul.execute({&projectionPrep, &inputPrep}, {&projected},  {}, {}, {});
+        mmul.execute({&projectionPrep, &inputPrep}, {&projected});
 
         projected.reshapei({numHeads, projectedSize, miniBatchSize, seqLength});
         projected.permutei({2, 0, 1, 3});   //[minibatch, numHeads, projectedSize, seqLength]
@@ -66,7 +66,7 @@ namespace nd4j {
         nd4j::ops::matmul_bp mmulBp;
         NDArray dLdProjectionPrep(projectionPrep.shapeInfo(), false, context);
         NDArray dLdInputPrep(inputPrep.shapeInfo(), false, context);
-        mmulBp.execute({&projectionPrep, &inputPrep, &epsReshaped}, {&dLdProjectionPrep, &dLdInputPrep}, {}, {}, {});
+        mmulBp.execute({&projectionPrep, &inputPrep, &epsReshaped}, std::vector<NDArray*>{&dLdProjectionPrep, &dLdInputPrep}, {}, {}, {});
 
         dLdProjectionPrep.reshapei({numHeads, projectionMatrix->sizeAt(1), projectionMatrix->sizeAt(2)});
         dLdProjectionMatrix->assign(dLdProjectionPrep);

libnd4j/include/helpers/impl/ShapeUtils.cpp

@@ -1019,15 +1019,6 @@ std::vector<int> ShapeUtils::tadAxesForSimpleBroadcast(const NDArray& max, const
     return numOfMinTads == 1 ? maxTadDims : std::vector<int>();
 }
 
-
-Nd4jLong ShapeUtils::stringBufferHeaderRequirements(Nd4jLong numStrings) {
-    // we store +1 offset
-    auto base = numStrings + 1;
-
-    // since we return number of bytes...
-    return base * sizeof(Nd4jLong);
-}
-
 ////////////////////////////////////////////////////////////////////////////////
 /*
 bool ShapeUtils::isSubArrayCase(const NDArray& arr1, const NDArray& arr2, std::vector<int>& sameDims) {

libnd4j/include/helpers/impl/StringUtils.cpp

@@ -1,5 +1,6 @@
 /*******************************************************************************
  * Copyright (c) 2015-2018 Skymind, Inc.
+ * Copyright (c) 2019-2020 Konduit K.K.
  *
  * This program and the accompanying materials are made available under the
  * terms of the Apache License, Version 2.0 which is available at
@@ -16,6 +17,7 @@
 
 //
 // Created by raver119 on 20/04/18.
+// @author Oleg Semeniv <oleg.semeniv@gmail.com>
 //
 
 #include <helpers/StringUtils.h>
@@ -49,13 +51,8 @@ namespace nd4j {
         if (!array.isS())
             throw nd4j::datatype_exception::build("StringUtils::byteLength expects one of String types;", array.dataType());
 
-        uint64_t result = 0;
-
-        // our buffer stores offsets, and the last value is basically number of bytes used
         auto buffer = array.bufferAsT<Nd4jLong>();
-        result = buffer[array.lengthOf()];
+        return buffer[array.lengthOf()];
-
-        return result;
     }
 
     std::vector<std::string> StringUtils::split(const std::string &haystack, const std::string &delimiter) {
@@ -73,4 +70,89 @@ namespace nd4j {
 
         return output;
     }
+    
+    bool StringUtils::u8StringToU16String(const std::string& u8, std::u16string& u16) {
+
+        if (u8.empty()) 
+            return false;
+
+        u16.resize(unicode::offsetUtf8StringInUtf16(u8.data(), u8.size()) / sizeof(char16_t));
+        if (u8.size() == u16.size()) 
+            u16.assign(u8.begin(), u8.end());
+        else
+            return unicode::utf8to16(u8.data(), &u16[0], u8.size());
+        
+        return true;
+    }
+
+    bool StringUtils::u8StringToU32String(const std::string& u8, std::u32string& u32) {
+
+        if (u8.empty()) 
+            return false;
+
+        u32.resize( unicode::offsetUtf8StringInUtf32(u8.data(), u8.size()) / sizeof(char32_t) );
+        if (u8.size() == u32.size()) 
+            u32.assign(u8.begin(), u8.end());
+        else 
+            return unicode::utf8to32(u8.data(), &u32[0], u8.size());
+        
+        return true;
+    }
+
+    bool StringUtils::u16StringToU32String(const std::u16string& u16, std::u32string& u32) {
+        
+        if (u16.empty()) 
+            return false;
+
+        u32.resize(unicode::offsetUtf16StringInUtf32(u16.data(), u16.size()) / sizeof(char32_t));
+        if (u16.size() == u32.size()) 
+            u32.assign(u16.begin(), u16.end());
+        else 
+            return unicode::utf16to32(u16.data(), &u32[0], u16.size());
+        
+        return true;
+    }
+    
+    bool StringUtils::u16StringToU8String(const std::u16string& u16, std::string& u8) {
+
+        if (u16.empty()) 
+            return false;
+        
+        u8.resize(unicode::offsetUtf16StringInUtf8(u16.data(), u16.size()));
+        if (u16.size() == u8.size()) 
+            u8.assign(u16.begin(), u16.end());
+        else
+            return unicode::utf16to8(u16.data(), &u8[0], u16.size());
+        
+        return true;
+    }
+    
+    bool StringUtils::u32StringToU16String(const std::u32string& u32, std::u16string& u16) {
+        
+        if (u32.empty()) 
+            return false;
+
+        u16.resize(unicode::offsetUtf32StringInUtf16(u32.data(), u32.size()) / sizeof(char16_t));
+        if (u32.size() == u16.size()) 
+            u16.assign(u32.begin(), u32.end());
+        else 
+            return unicode::utf32to16(u32.data(), &u16[0], u32.size());
+        
+        return true;
+    }
+    
+    bool StringUtils::u32StringToU8String(const std::u32string& u32, std::string& u8) {
+        
+        if (u32.empty()) 
+            return false;
+
+        u8.resize(unicode::offsetUtf32StringInUtf8(u32.data(), u32.size()));
+        if (u32.size() == u8.size()) 
+            u8.assign(u32.begin(), u32.end());
+        else 
+            return unicode::utf32to8(u32.data(), &u8[0], u32.size());
+
+        return true;
+    }
+
 }

libnd4j/include/helpers/impl/unicode.cpp

@@ -0,0 +1,456 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2020 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 Oleg Semeniv <oleg.semeniv@gmail.com>
+//
+
+#include <unicode.h>
+
+namespace nd4j {
+namespace unicode {
+
+    constexpr uint32_t ONEBYTEBOUND = 0x00000080;
+    constexpr uint32_t TWOBYTEBOUND = 0x00000800;
+    constexpr uint32_t THREEBYTEBOUND = 0x00010000;
+    constexpr uint16_t HIGHBYTEMIN = 0xd800u;
+    constexpr uint16_t HIGHBYTEMAX = 0xdbffu;
+    constexpr uint16_t TRAILBYTEMIN = 0xdc00u;
+    constexpr uint16_t TRAILBYTEMAX = 0xdfffu;
+    constexpr uint16_t HIGHBYTEOFFSET = HIGHBYTEMIN - (0x10000 >> 10);
+    constexpr uint32_t BYTEOFFSET = 0x10000u - (HIGHBYTEMIN << 10) - TRAILBYTEMIN;
+    // Maximum valid value for a Unicode code point
+    constexpr uint32_t CODEPOINTMAX = 0x0010ffffu;
+
+    template<typename T>
+    FORCEINLINE uint8_t castToU8(const T cp) {
+        return static_cast<uint8_t>(0xff & cp);
+    }
+
+    template<typename T>
+    FORCEINLINE uint16_t castToU16(const T cp) {
+        return static_cast<uint16_t>(0xffff & cp);
+    }
+
+    template<typename T>
+    FORCEINLINE uint32_t castToU32(const T cp) {
+        return static_cast<uint32_t>(0xffffff & cp);
+    }
+
+    template<typename T>
+    FORCEINLINE bool isTrail(const T cp) {
+        return ((castToU8(cp) >> 6) == 0x2);
+    }
+
+    template <typename T>
+    FORCEINLINE bool isHighSurrogate(const T cp) {
+        return (cp & 0xfffffc00) == 0xd800;
+    }
+
+    template <typename T>
+    bool isLowSurrogate(const T cp) {
+        return (cp & 0xfffffc00) == 0xdc00;
+    }
+
+    template <typename T>
+    FORCEINLINE bool isLeadSurrogate(const T cp) {
+        return (cp >= HIGHBYTEMIN && cp <= HIGHBYTEMAX);
+    }
+
+    template <typename T>
+    FORCEINLINE bool isTrailSurrogate(const T cp) {
+        return (cp >= TRAILBYTEMIN && cp <= TRAILBYTEMAX);
+    }
+
+    template <typename T>
+    FORCEINLINE bool isSurrogateU8(const T cp) {
+        return (cp >= HIGHBYTEMIN && cp <= TRAILBYTEMAX);
+    }
+
+    template <typename T>
+    FORCEINLINE bool isSurrogateU16(const T cp) {
+        return ((cp - 0xd800u) < 2048u);
+    }
+
+    template <typename T>
+    FORCEINLINE bool isSymbolU8Valid(const T cp) {
+        return (cp <= CODEPOINTMAX && !isSurrogateU8(cp));
+    }
+
+    template <typename T>
+    FORCEINLINE bool isSymbolValid(const T cp) {
+        return (cp <= CODEPOINTMAX);
+    }
+
+    template <typename T>
+    FORCEINLINE uint32_t surrogateU32(const T& high, const T& low) {
+        return (high << 10) + low - 0x35fdc00;
+    }
+
+    template <typename T>
+    Nd4jLong symbolLength(const T* it) {
+        uint8_t lead = castToU8(*it);
+        if (lead < 0x80)
+            return 1;
+        else if ((lead >> 5) == 0x6)
+            return 2;
+        else if ((lead >> 4) == 0xe)
+            return 3;
+        else if ((lead >> 3) == 0x1e)
+            return 4;
+        else
+            return 0;
+    }
+
+    template <typename T>
+    Nd4jLong symbolLength32(const T* it) {
+        auto lead = castToU32(*it);
+        if (lead < ONEBYTEBOUND)
+            return 1;
+        else if (lead < TWOBYTEBOUND)
+            return 2;
+        else if (lead < THREEBYTEBOUND)
+            return 3;
+        else if (lead <= CODEPOINTMAX)
+            return 4;
+        else
+            return 0;
+    }
+
+    template <typename T>
+    Nd4jLong symbolLength16(const T* it) {
+
+        uint32_t lead = castToU16(*it);
+        if (!isLeadSurrogate(lead)) {
+            if (lead < ONEBYTEBOUND)
+                return 1;
+            else if (lead < TWOBYTEBOUND)
+                return 2;
+            else if (lead < THREEBYTEBOUND)
+                return 3;
+            else
+                return 0;
+        }
+        else {
+            return 4;
+        }
+    }
+
+    Nd4jLong offsetUtf8StringInUtf32(const void* start, const void* end) {
+        
+        Nd4jLong count = 0;
+        for (auto it = static_cast<const int8_t*>(start); it != end; it++) {
+            auto length = symbolLength(it);
+            it += (length > 0) ? (length - 1) : 0;
+            count += 1;
+        }
+        return static_cast<Nd4jLong>(count * sizeof(char32_t));
+    }
+    
+    Nd4jLong offsetUtf16StringInUtf32(const void* start, const void* end) {
+
+        Nd4jLong count = 0;
+        for (auto it = static_cast<const uint16_t*>(start); it != end;) {
+            auto length = symbolLength16(it);
+            it += (4 == length) ? 2 : 1;
+            count += 1;
+        }
+        return static_cast<Nd4jLong>(count*sizeof(char32_t));
+    }
+    
+    Nd4jLong offsetUtf8StringInUtf16(const void* start, const void* end) {
+
+        Nd4jLong count = 0;
+        for (auto it = static_cast<const int8_t*>(start); it != end; it++) {
+            auto length = symbolLength(it);
+            auto step = ((length > 0) ? (length - 1) : 0);
+            it += step;
+            count += (4 == length) ? 2 : 1;
+        }
+        return static_cast<Nd4jLong>(count*sizeof(char16_t));
+    }
+    
+    Nd4jLong offsetUtf16StringInUtf8(const void* start, const void* end) {
+
+        Nd4jLong count = 0;
+        for (auto it = static_cast<const uint16_t*>(start); it != end;) {
+            auto length = symbolLength16(it);
+            it += (4 == length) ? 2 : 1;
+            count += length;
+        }
+        return static_cast<Nd4jLong>(count);
+    }
+    
+    Nd4jLong offsetUtf32StringInUtf16(const void* start, const void* end) {
+        
+        Nd4jLong count = 0;
+        for (auto it = static_cast<const uint32_t*>(start); it != end; it++) {
+            auto length = symbolLength32(it);
+            count += (4 == length) ? 2 : 1;;
+        }
+        return static_cast<Nd4jLong>(count*sizeof(char16_t));
+    }
+    
+    Nd4jLong offsetUtf32StringInUtf8(const void* start, const void* end) {
+
+        Nd4jLong count = 0;
+        for (auto it = static_cast<const uint32_t*>(start); it != end; it++) {
+            count += symbolLength32(it);
+        }
+        return count;
+    }
+    
+    bool isStringValidU8(const void* start, const void* stop) {
+        for (auto it = static_cast<const int8_t*>(start); it != stop; it++) {
+            if (!isSymbolU8Valid( castToU8(*it) )) {
+                return false;
+            }
+        }
+        return true;
+    }
+    
+    bool isStringValidU16(const void* start, const void* stop) {
+        for (auto it = static_cast<const uint16_t*>(start); it != stop; it++) {
+            if (!isSymbolValid( castToU32(*it) )) {
+                return false;
+            }
+        }
+        return true;
+    }
+    
+    bool isStringValidU32(const void* start, const void* stop) {
+        for (auto it = static_cast<const uint32_t*>(start); it != stop; it++) {
+            if (!isSymbolValid( castToU32(*it) )) {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    void* utf16to8Ptr(const void* start, const void* end, void* res) {
+
+        auto result = static_cast<int8_t*>(res);
+        // result have to be  pre-allocated
+        for (auto it = static_cast<const uint16_t*>(start); it != end;) {
+            uint32_t cp = castToU16(*it++);
+             if (!isLeadSurrogate(cp)) {
+                 if (cp < 0x80) {                        // for one byte
+                     *(result++) = static_cast<uint8_t>(cp);
+                 }
+                 else if (cp < 0x800) {                // for two bytes
+                     *(result++) = static_cast<uint8_t>((cp >> 6) | 0xc0);
+                     *(result++) = static_cast<uint8_t>((cp & 0x3f) | 0x80);
+                 }
+                 else{              // for three bytes
+                     *(result++) = static_cast<uint8_t>((cp >> 12) | 0xe0);
+                     *(result++) = static_cast<uint8_t>(((cp >> 6) & 0x3f) | 0x80);
+                     *(result++) = static_cast<uint8_t>((cp & 0x3f) | 0x80);
+                 }
+             }
+             else {
+                 if (it != end) {
+                     uint32_t trail_surrogate = castToU16(*it++);
+                     if (isTrailSurrogate(trail_surrogate))
+                         cp = (cp << 10) + trail_surrogate + BYTEOFFSET;
+                 }
+                  // for four bytes
+                 *(result++) = static_cast<uint8_t>((cp >> 18) | 0xf0);
+                 *(result++) = static_cast<uint8_t>(((cp >> 12) & 0x3f) | 0x80);
+                 *(result++) = static_cast<uint8_t>(((cp >> 6) & 0x3f) | 0x80);
+                 *(result++) = static_cast<uint8_t>((cp & 0x3f) | 0x80);
+             }
+         }
+         return result;
+     }
+     
+     void* utf8to16Ptr(const void* start, const void* end, void* res) {
+         
+         auto result = static_cast<uint16_t*>(res);
+         // result have to be  pre-allocated
+         for (auto it = static_cast<const int8_t*>(start); it != end;) {
+             
+             auto nLength = symbolLength(it);
+             uint32_t cp = castToU8(*it++);
+             if (4 != nLength) {
+                 if (2 == nLength) {
+                     cp = ((cp << 6) & 0x7ff) + ((*it++) & 0x3f);
+                 }
+                 else if (3 == nLength) {
+                     cp = ((cp << 12) & 0xffff) + ((castToU8(*it++) << 6) & 0xfff);
+                     cp += (*it++) & 0x3f;
+                 }
+                 *(result++) = static_cast<uint16_t>(cp);
+             }
+             else {
+                 cp = ((cp << 18) & 0x1fffff) + ((castToU8(*it++) << 12) & 0x3ffff);
+                 cp += (castToU8(*it++) << 6) & 0xfff;
+                 cp += (*it++) & 0x3f;
+                 //make a surrogate pair
+                 *(result++) = static_cast<uint16_t>((cp >> 10) + HIGHBYTEOFFSET);
+                 *(result++) = static_cast<uint16_t>((cp & 0x3ff) + TRAILBYTEMIN);
+             }
+         }
+         return result;
+     }
+
+     void* utf32to8Ptr( const void* start, const void* end, void* result) {
+          
+         auto res = static_cast<uint8_t*>(result);
+         // result have to be  pre-allocated
+         for (auto it = static_cast<const uint32_t*>(start); it != end; it++) {
+
+            if (*it < 0x80)                        // for one byte
+                 *(res++) = static_cast<uint8_t>(*it);
+             else if (*it < 0x800) {                // for two bytes
+                 *(res++) = static_cast<uint8_t>((*it >> 6) | 0xc0);
+                 *(res++) = static_cast<uint8_t>((*it & 0x3f) | 0x80);
+             }
+             else if (*it < 0x10000) {              // for three bytes
+                 *(res++) = static_cast<uint8_t>((*it >> 12) | 0xe0);
+                 *(res++) = static_cast<uint8_t>(((*it >> 6) & 0x3f) | 0x80);
+                 *(res++) = static_cast<uint8_t>((*it & 0x3f) | 0x80);
+             }
+             else {                                // for four bytes
+                 *(res++) = static_cast<uint8_t>((*it >> 18) | 0xf0);
+                 *(res++) = static_cast<uint8_t>(((*it >> 12) & 0x3f) | 0x80);
+                 *(res++) = static_cast<uint8_t>(((*it >> 6) & 0x3f) | 0x80);
+                 *(res++) = static_cast<uint8_t>((*it & 0x3f) | 0x80);
+             }
+         }
+         return result;
+     }
+
+     void* utf8to32Ptr(const void* start, const void* end, void* res) {
+         
+         auto result = static_cast<uint32_t*>(res);
+         // result have to be  pre-allocated
+         for (auto it = static_cast<const int8_t*>(start); it != end;) {
+             
+             auto nLength = symbolLength(it);
+             uint32_t cp = castToU8(*it++);
+             if (2 == nLength) {
+                 cp = ((cp << 6) & 0x7ff) + ((*it++) & 0x3f);
+             }
+             else if (3 == nLength) {
+                 cp = ((cp << 12) & 0xffff) + ((castToU8(*it++) << 6) & 0xfff);
+                 cp += (*it++) & 0x3f;
+             }
+             else if (4 == nLength) {
+                 cp = ((cp << 18) & 0x1fffff) + ((castToU8(*it++) << 12) & 0x3ffff);
+                 cp += (castToU8(*it++) << 6) & 0xfff;
+                 cp += (*it++) & 0x3f;
+             }
+             (*result++) = cp;
+         }
+         return result;
+     }
+    
+     void* utf16to32Ptr(const void* start, const void* end, void* res) {
+
+         auto result = static_cast<uint32_t*>(res);
+         // result have to be  pre-allocated
+         for (auto it = static_cast<const uint16_t*>(start); it != end; it++) {
+             
+             uint32_t cpHigh = castToU32(*it);
+             if (!isSurrogateU16(cpHigh)) {
+                 *result++ = cpHigh;
+             }
+             else {
+                 it++;
+                 uint32_t cpLow = castToU32(*it);
+                 if (isHighSurrogate(cpHigh) && it != end && isLowSurrogate(cpLow)) {
+                     *result++ = surrogateU32(cpHigh, cpLow);
+                 }
+             }
+         }
+         return result;
+     }
+
+     void* utf32to16Ptr(const void* start, const void* end, void* res) {
+         
+         auto result = static_cast<uint16_t*>(res);
+         // result have to be  pre-allocate
+         for (auto it = static_cast<const uint32_t*>(start); it != end; it++) {
+
+             uint32_t cpHigh = castToU32(*it);
+             // todo check do we need this as we have pre-validation, if yes find out how to check u16
+             if (cpHigh < 0 || cpHigh > 0x10FFFF || (cpHigh >= 0xD800 && cpHigh <= 0xDFFF)) {
+                 // Invalid code point.  Replace with sentinel, per Unicode standard:
+                 *result++ = u'\uFFFD';
+             }
+             else if (cpHigh < 0x10000UL) { // In the BMP.
+                 *result++ = static_cast<char16_t>(cpHigh);
+             }
+             else {
+                 *result++ = static_cast<char16_t>(((cpHigh - 0x10000UL) / 0x400U) + 0xD800U);
+                 *result++ = static_cast<char16_t>(((cpHigh - 0x10000UL) % 0x400U) + 0xDC00U);
+             }
+         }
+         return result;
+     }
+
+     Nd4jLong offsetUtf8StringInUtf32(const void* input, uint32_t nInputSize) {
+         return  offsetUtf8StringInUtf32(input, static_cast<const int8_t*>(input) + nInputSize);
+     }
+
+     Nd4jLong offsetUtf16StringInUtf32(const void* input, uint32_t nInputSize) {
+         return  offsetUtf16StringInUtf32(input, static_cast<const uint16_t*>(input) + nInputSize);
+     }
+    
+     Nd4jLong offsetUtf8StringInUtf16(const void* input, uint32_t nInputSize) {
+         return offsetUtf8StringInUtf16(input, static_cast<const int8_t*>(input) + nInputSize);
+     }
+ 
+     Nd4jLong offsetUtf16StringInUtf8(const void* input, uint32_t nInputSize) {
+         return offsetUtf16StringInUtf8(input, static_cast<const uint16_t*>(input) + nInputSize);
+     }
+
+     Nd4jLong offsetUtf32StringInUtf8(const void* input, uint32_t nInputSize) {
+         return offsetUtf32StringInUtf8(input, static_cast<const uint32_t*>(input) + nInputSize);
+     }
+     
+     Nd4jLong offsetUtf32StringInUtf16(const void* input, const uint32_t nInputSize) {
+         return offsetUtf32StringInUtf16(input, static_cast<const uint32_t*>(input) + nInputSize);
+     }
+     
+     bool utf8to16(const void* input, void* output, uint32_t nInputSize) {
+         return utf8to16Ptr(input, static_cast<const int8_t*>(input) + nInputSize, output);
+     }
+     
+     bool utf8to32(const void* input, void* output, uint32_t nInputSize) {
+         return utf8to32Ptr(input, static_cast<const int8_t*>(input) + nInputSize, output);
+     }
+     
+     bool utf16to32(const void* input, void* output, uint32_t nInputSize) {
+         return utf16to32Ptr(input, static_cast<const uint16_t*>(input) + nInputSize, output);
+     }
+     
+     bool utf16to8(const void* input, void* output, uint32_t nInputSize) {
+         return utf16to8Ptr(input, static_cast<const uint16_t*>(input) + nInputSize, output);
+     }
+     
+     bool utf32to16(const void* input, void* output, uint32_t nInputSize) {
+         return utf32to16Ptr(input, static_cast<const uint32_t*>(input) + nInputSize, output);
+     }
+     
+     bool utf32to8(const void* input, void* output, const Nd4jLong nInputSize) {
+         return utf32to8Ptr(input, static_cast<const uint32_t*>(input) + nInputSize, output);
+     }
+
+ }
+
+}
+

libnd4j/include/helpers/unicode.h

@@ -0,0 +1,189 @@
+/*******************************************************************************
+ * Copyright (c) 2019-2020 Konduit K.K.
+ *
+ * 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 Oleg Semeniv <oleg.semeniv@gmail.com>
+//
+
+#ifndef LIBND4J_UNICODE_H
+#define LIBND4J_UNICODE_H
+
+#include <NDArray.h>
+
+namespace nd4j {
+namespace unicode {
+
+    /**
+        * This method calculate u16 offset based on utf8
+        * @param const pointer to the utf8 string start point
+        * @param size of the string
+        * @return offset of utf16
+   */
+    Nd4jLong offsetUtf8StringInUtf16(const void* start, const void* end);
+
+    /**
+        * This method calculate u8 offset based on utf16
+        * @param const pointer to the utf16 string start point
+        * @param size of the string
+        * @return offset of utf8
+   */
+    Nd4jLong offsetUtf16StringInUtf8(const void* start, const void* end);
+
+    /**
+       * This method calculate u32 offset based on utf16
+       * @param const pointer to the utf16 string start point
+       * @param size of the string
+       * @return offset of utf32
+    */
+    Nd4jLong offsetUtf32StringInUtf16(const void* start, const void* end);
+
+    /**
+       * This method calculate u32 offset based on utf8
+       * @param const pointer to the utf16 string start point
+       * @param size of the string
+       * @return offset of utf8
+    */
+    Nd4jLong offsetUtf32StringInUtf8(const void* start, const void* end);
+
+    /*
+    * This function check is valid charecter in u8 string
+    */
+    bool isStringValidU8(const void* start, const void* stop);
+
+    /*
+    * This function check is valid charecter in u16 string
+    */
+    bool isStringValidU16(const void* start, const void* stop);
+
+    /*
+    * This function check is valid u32 charecter in string
+    */
+    bool isStringValidU32(const void* start, const void* stop);
+
+    /**
+         * This method count offset for utf8 string in utf32
+         * @param const pointer to the utf8 string start point
+         * @param size of the string
+         * @return offset
+    */
+    Nd4jLong offsetUtf8StringInUtf32(const void* input, uint32_t nInputSize);
+
+    /**
+     * This method count offset for utf8 string in utf32
+     * @param const pointer to the utf8 string start point
+     * @param const end pointer to the utf8 string
+     * @return offset
+    */
+    Nd4jLong offsetUtf8StringInUtf32(const void* input, const void* stop);
+
+    /**
+         * This method count offset for utf32 based on utf16 string
+         * @param const pointer to the utf16 string start point
+         * @param size of the string
+         * @return offset
+    */
+    Nd4jLong offsetUtf16StringInUtf32(const void* input, uint32_t nInputSize);
+
+    /**
+         * This method calculate offset of u16 based on utf8
+         * @param const pointer to the utf8 string start point
+         * @param size of the string
+         * @return offset of utf16
+    */
+    Nd4jLong offsetUtf8StringInUtf16(const void* input, uint32_t nInputSize);
+
+    /**
+        * This method calculate offset of u8 based on utf16
+        * @param const pointer to the utf16 string start point
+        * @param size of the string
+        * @return offset of utf8
+   */
+    Nd4jLong offsetUtf16StringInUtf8(const void* input, uint32_t nInputSize);
+
+    /**
+       * This method calculate offset of u32 based on utf8
+       * @param const pointer to the utf16 string start point
+       * @param size of the string
+       * @return offset of utf32
+    */
+    Nd4jLong offsetUtf32StringInUtf8(const void* input, uint32_t nInputSize);
+
+    /**
+       * This method calculate offset of u32 based on utf16
+       * @param const pointer to the utf16 string start point
+       * @param size of the string
+       * @return offset of utf32
+    */
+    Nd4jLong offsetUtf32StringInUtf16(const void* input, const uint32_t nInputSize);
+
+    /**
+         * This method convert utf8 string to utf16 string
+         * @param const pointer to the utf8 string start point
+         * @param reference to start point to utf16 
+         * @param size of input utf8 string
+         * @return status of convertion
+    */
+    bool utf8to16(const void* input, void* output, uint32_t nInputSize);
+
+    /**
+         * This method convert utf8 string to utf32 string
+         * @param const pointer to the utf8 string start point
+         * @param reference to start point to utf32
+         * @param size of input utf8 string
+         * @return status of convertion
+    */
+    bool utf8to32(const void* input, void* output, uint32_t nInputSize);
+
+    /**
+         * This method convert utf16 string to utf32 string
+         * @param const pointer to the utf16 string start point
+         * @param reference to start point to utf32
+         * @param size of input utf16 string
+         * @return status of convertion
+    */
+    bool utf16to32(const void* input, void* output, uint32_t nInputSize);
+
+    /**
+         * This method convert utf16 string to utf8 string
+         * @param const pointer to the utf16 string start point
+         * @param reference to start point to utf8
+         * @param size of input utf16 string
+         * @return status of convertion
+    */
+    bool utf16to8(const void* input, void* output, uint32_t nInputSize);
+
+    /**
+         * This method convert utf32 string to utf16 string
+         * @param const pointer to the utf32 string start point
+         * @param reference to start point to utf16
+         * @param size of input utf32 string
+         * @return status of convertion
+    */
+    bool utf32to16(const void* input, void* output, uint32_t nInputSize);
+
+    /**
+         * This method convert utf32 string to utf8 string
+         * @param const pointer to the utf32 string start point
+         * @param reference to start point to utf8
+         * @param size of input utf32 string
+         * @return status of convertion
+    */
+    bool utf32to8(const void* input, void* output, const Nd4jLong nInputSize);
+}
+}
+
+
+#endif //LIBND4J_UNICODE_H

libnd4j/include/loops/cpu/random.cpp

@@ -29,6 +29,7 @@ using namespace randomOps;
 namespace functions {
     namespace random {
 
+
         template<typename X>
         template<typename OpClass>
         void RandomFunction<X>::execTransform(Nd4jPointer state,
@@ -56,18 +57,32 @@ namespace functions {
 
             if(shape::haveSameShapeAndStrides(xShapeInfo, yShapeInfo) && shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo)) {
 
-                uint xShapeInfoCast[MAX_RANK];
-                const bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
 
-                auto func = PRAGMA_THREADS_FOR {
+                if(shape::elementWiseStride(zShapeInfo) ==  1 && shape::elementWiseStride(xShapeInfo) ==  1 && shape::elementWiseStride(yShapeInfo) ==  1 &&
-                    PRAGMA_OMP_SIMD
+                        shape::order(xShapeInfo) == shape::order(zShapeInfo) && shape::order(zShapeInfo) == shape::order(yShapeInfo) ){
-                    for (auto i = start; i < stop; i += increment)  {
-                        auto offset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
-                        z[offset] = OpClass::op(x[offset], y[offset], i, length, rng, extraArguments);
-                    }
-                };
                     
-                samediff::Threads::parallel_for(func,  0, length, 1);
+                    auto func = PRAGMA_THREADS_FOR {
+                        PRAGMA_OMP_SIMD
+                        for (auto i = start; i < stop; i++)  { 
+                            z[i] = OpClass::op(x[i], y[i], i, length, rng, extraArguments);
+                        }
+                    };
+                    samediff::Threads::parallel_for(func,  0, length, 1);
+                }
+                else{
+                    uint xShapeInfoCast[MAX_RANK];
+                    const bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+
+                    auto func = PRAGMA_THREADS_FOR {
+                        PRAGMA_OMP_SIMD
+                        for (auto i = start; i < stop; i += increment)  {
+                            auto offset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
+                            z[offset] = OpClass::op(x[offset], y[offset], i, length, rng, extraArguments);
+                        }
+                    };
+
+                    samediff::Threads::parallel_for(func,  0, length, 1);
+                }
             }
             else if (shape::haveSameShapeAndStrides(xShapeInfo, yShapeInfo)) {
 
@@ -169,15 +184,27 @@ namespace functions {
 
             if(shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo)) {
 
-                auto func = PRAGMA_THREADS_FOR {
+                if(shape::elementWiseStride(zShapeInfo) ==  1 &&  shape::elementWiseStride(xShapeInfo) ==  1 && shape::order(xShapeInfo) == shape::order(zShapeInfo)){
-                    PRAGMA_OMP_SIMD
-                    for (uint64_t i = start; i < stop; i += increment)  {
-                        auto offset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
-                        z[offset] = OpClass::op(x[offset], i, length, rng, extraArguments);
-                    }
-                };
                     
-                samediff::Threads::parallel_for(func,  0, length, 1);
+                    auto func = PRAGMA_THREADS_FOR {
+                        PRAGMA_OMP_SIMD
+                        for (auto i = start; i < stop; i++)  { 
+                            z[i] = OpClass::op(x[i], i, length, rng, extraArguments);
+                        }
+                    };
+                    samediff::Threads::parallel_for(func,  0, length, 1);
+                }
+                else{
+                    auto func = PRAGMA_THREADS_FOR {
+                        PRAGMA_OMP_SIMD
+                        for (uint64_t i = start; i < stop; i += increment)  {
+                            auto offset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
+                            z[offset] = OpClass::op(x[offset], i, length, rng, extraArguments);
+                        }
+                    };
+
+                    samediff::Threads::parallel_for(func,  0, length, 1);
+                }
             }
             else {
 
@@ -208,20 +235,34 @@ namespace functions {
             auto length = shape::length(zShapeInfo);
 
             nd4j::graph::RandomGenerator* rng = reinterpret_cast<nd4j::graph::RandomGenerator*>(state);
-            nd4j::OmpLaunchHelper info(length);
 
-            uint zShapeInfoCast[MAX_RANK];
+            if(shape::elementWiseStride(zShapeInfo) ==  1){
-            const bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zShapeInfo, zShapeInfoCast);
 
-            auto func = PRAGMA_THREADS_FOR {
+                auto func = PRAGMA_THREADS_FOR {
-                PRAGMA_OMP_SIMD
+                        PRAGMA_OMP_SIMD
-                for (uint64_t i = start; i < stop; i += increment)  {
+                        for (auto i = start; i < stop; i++)  { 
-                    auto offset = shape::indexOffset(i, zShapeInfo, zShapeInfoCast, canCastZ);
+                            z[i] = OpClass::op( i, length, rng, extraArguments);
-                    z[offset] = OpClass::op(i, length, rng, extraArguments);
+                        }
-                }
+                };
-            };
 
-            samediff::Threads::parallel_for(func,  0, length, 1);
+                samediff::Threads::parallel_for(func,  0, length, 1); 
+            }
+            else{
+                nd4j::OmpLaunchHelper info(length);
+
+                uint zShapeInfoCast[MAX_RANK];
+                const bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zShapeInfo, zShapeInfoCast);
+
+                auto func = PRAGMA_THREADS_FOR {
+                    PRAGMA_OMP_SIMD
+                    for (uint64_t i = start; i < stop; i += increment)  {
+                        auto offset = shape::indexOffset(i, zShapeInfo, zShapeInfoCast, canCastZ);
+                        z[offset] = OpClass::op(i, length, rng, extraArguments);
+                    }
+                };
+
+                samediff::Threads::parallel_for(func,  0, length, 1);
+            }
         }
 
         template<typename X>

libnd4j/include/op_boilerplate.h

@@ -1516,7 +1516,9 @@
 
 #define INPUT_LIST(INDEX)     reinterpret_cast<nd4j::NDArrayList *>(block.getVariable(INDEX)->getNDArrayList())
 
+#define D_ARG(INDEX)     block.getDArguments()->at(INDEX)
 #define INT_ARG(INDEX)     block.getIArguments()->at(INDEX)
+#define I_ARG(INDEX)     INT_ARG(INDEX)
 #define T_ARG(INDEX)     block.getTArguments()->at(INDEX)
 #define B_ARG(INDEX)     block.getBArguments()->at(INDEX)
 

libnd4j/include/ops/declarable/BooleanOp.h

@@ -36,9 +36,8 @@ namespace nd4j {
         public:
             BooleanOp(const char *name, int numInputs, bool scalar);
 
-            bool evaluate(std::initializer_list<nd4j::NDArray*> args);
+            bool verify(const std::vector<nd4j::NDArray*>& args);
-            bool evaluate(std::vector<nd4j::NDArray*>& args);
+            bool verify(nd4j::graph::Context& block);
-            bool evaluate(nd4j::graph::Context& block);
 
             Nd4jStatus execute(Context* block) override;
 

libnd4j/include/ops/declarable/DeclarableOp.h

@@ -169,13 +169,22 @@ namespace nd4j {
              */
             virtual Nd4jStatus execute(Context* block);
 
-            nd4j::ResultSet* execute(std::initializer_list<NDArray*> inputs, std::initializer_list<double> tArgs, std::initializer_list<Nd4jLong> iArgs, std::initializer_list<bool> bArgs, bool isInplace = false, nd4j::DataType type = nd4j::DataType::FLOAT32);
+            Nd4jStatus execute(const std::vector<NDArray*> &inputs, const std::vector<NDArray*> &outputs);
-            Nd4jStatus execute(std::initializer_list<NDArray*> inputs, std::initializer_list<NDArray*> outputs , std::initializer_list<double> tArgs, std::initializer_list<Nd4jLong> iArgs, std::initializer_list<bool> bArgs, bool isInplace = false, nd4j::DataType type = nd4j::DataType::FLOAT32);
-            Nd4jStatus execute(nd4j::graph::RandomGenerator& rng, std::initializer_list<NDArray*> inputs, std::initializer_list<NDArray*> outputs , std::initializer_list<double> tArgs, std::initializer_list<Nd4jLong> iArgs, std::initializer_list<bool> bArgs, bool isInplace = false, nd4j::DataType type = nd4j::DataType::FLOAT32);
 
-            nd4j::ResultSet* execute(const std::vector<NDArray*>& inputs, const std::vector<double>& tArgs, const std::vector<Nd4jLong>& iArgs, const std::vector<bool>& bArgs = std::vector<bool>(), bool isInplace = false, nd4j::DataType type = nd4j::DataType::FLOAT32);
+            template <class T, typename = std::enable_if<DataTypeUtils::scalarTypesForExecution<T>::value>>
-            Nd4jStatus execute(std::vector<NDArray*>& inputs, std::vector<NDArray*>& outputs , std::vector<double>& tArgs, std::vector<Nd4jLong>& iArgs, std::vector<bool>& bArgs, bool isInplace = false, nd4j::DataType type = nd4j::DataType::FLOAT32);
+            Nd4jStatus execute(const std::vector<NDArray*> &inputs, const std::vector<NDArray*> &outputs, std::initializer_list<T> tArgs);
-            Nd4jStatus execute(nd4j::graph::RandomGenerator& rng, std::vector<NDArray*>& inputs, std::vector<NDArray*>& outputs, std::vector<double>& tArgs, std::vector<Nd4jLong>& iArgs, std::vector<bool>& bArgs, bool isInplace = false, nd4j::DataType type = nd4j::DataType::FLOAT32);
+
+            Nd4jStatus execute(const std::vector<NDArray*> &inputs, const std::vector<NDArray*> &outputs, const std::vector<double> &tArgs, const std::vector<Nd4jLong> &iArgs, const std::vector<bool> &bArgs = std::vector<bool>(), const std::vector<nd4j::DataType> &dArgs = std::vector<nd4j::DataType>(), bool isInplace = false);
+
+
+            nd4j::ResultSet* evaluate(const std::vector<NDArray*> &inputs);
+
+            template <class T, typename = std::enable_if<DataTypeUtils::scalarTypesForExecution<T>::value>>
+            nd4j::ResultSet* evaluate(const std::vector<NDArray*> &inputs, std::initializer_list<T> args);
+
+            nd4j::ResultSet* evaluate(const std::vector<NDArray*> &inputs, const std::vector<double> &tArgs, const std::vector<Nd4jLong> &iArgs, const std::vector<bool> &bArgs = std::vector<bool>(), const std::vector<nd4j::DataType> &dArgs = std::vector<nd4j::DataType>(), bool isInplace = false);
+
+            Nd4jStatus execute(nd4j::graph::RandomGenerator& rng, const std::vector<NDArray*>& inputs, const std::vector<NDArray*>& outputs, const std::vector<double>& tArgs, const std::vector<Nd4jLong>& iArgs, const std::vector<bool>& bArgs, const std::vector<nd4j::DataType> &dArgs = std::vector<nd4j::DataType>(), bool isInplace = false, nd4j::DataType type = nd4j::DataType::FLOAT32);
 
             nd4j::ResultSet* execute(const nd4j::OpArgsHolder& holder, bool isInplace = false);
 

libnd4j/include/ops/declarable/generic/activations/crelu.cpp

@@ -73,7 +73,7 @@ namespace nd4j {
 
             // at first step we build fwd activation
             nd4j::ops::crelu op;
-            auto tmpResult = op.execute({input}, {}, {}, {});
+            auto tmpResult = op.evaluate({input});
             if (tmpResult->status() != ND4J_STATUS_OK)
                 return tmpResult->status();
 
@@ -84,7 +84,7 @@ namespace nd4j {
             helpers::reluDerivative(block.launchContext(), actv, epsilonNext);
             // now we split updated array into 2 chunks along last dimension
             nd4j::ops::concat_bp opc;
-            auto dec = opc.execute({input, input, actv}, {}, {-1}, {});
+            auto dec = opc.evaluate({input, input, actv}, {-1});
             if (dec->status() != ND4J_STATUS_OK)
                 return dec->status();
 

libnd4j/include/ops/declarable/generic/boolean/where_np.cpp

@@ -103,7 +103,7 @@ namespace nd4j {
 //                if (output->isEmpty())
                 Nd4jLong width = condition->rankOf();
                 nd4j::ops::Where op;
-                std::unique_ptr<ResultSet> res(op.execute({condition}, {}, {}, {}));
+                std::unique_ptr<ResultSet> res(op.evaluate({condition}));
                 REQUIRE_OK(res->status());
                 NDArray* whereTrue = res->at(0);
                 if (whereTrue->isEmpty())

libnd4j/include/ops/declarable/generic/broadcastable/floormod.cpp

@@ -66,7 +66,7 @@ namespace nd4j {
             auto gradY = OUTPUT_VARIABLE(1);
             gradX->assign(epsNext);
             nd4j::ops::floormod op;
-            std::unique_ptr<ResultSet> tmpResult(op.execute({x, y}, {}, {}, {}));
+            std::unique_ptr<ResultSet> tmpResult(op.evaluate({x, y}));
 
             if (gradY->rankOf() == gradX->rankOf())
                 epsNext->applyPairwiseTransform(pairwise::Multiply, *tmpResult->at(0), *gradY);

libnd4j/include/ops/declarable/generic/broadcastable/pow.cpp

@@ -118,7 +118,7 @@ namespace ops {
        DECLARE_TYPES(Pow_bp) {
            getOpDescriptor()
                ->setAllowedInputTypes({ ALL_FLOATS, ALL_INTS })
-               ->setAllowedOutputTypes({ ALL_FLOATS }); // TODO maybe wourth to add ALL_INTS
+               ->setAllowedOutputTypes({ ALL_FLOATS }); 
        }
 
 }

libnd4j/include/ops/declarable/generic/compat/compat_string_split.cpp

@@ -81,7 +81,7 @@ namespace nd4j {
             }
 
             // now once we have all strings in single vector time to fill
-            auto tmp = NDArrayFactory::string('c', {(Nd4jLong) strings.size()}, strings);
+            auto tmp = NDArrayFactory::string({(Nd4jLong) strings.size()}, strings);
             auto blen = StringUtils::byteLength(tmp) + ShapeUtils::stringBufferHeaderRequirements(strings.size());
 
             // for CUDA mostly

libnd4j/include/ops/declarable/generic/nn/batchnorm.cpp

@@ -197,8 +197,7 @@ CUSTOM_OP_IMPL(batchnorm_bp, 4, 3, false, 1, 2) {
     // ***** calculations ***** //
 
     // notations:
-    // f = g * (gamma * ((x - m) / (v + eps)^0.5) + beta) -> means dLdO * ff_output
+    // f = g * (gamma * ((x - m) / (v + eps)^0.5) + beta) -> means dLdO * ff_output, g = dLdO
-    // g = dLdO
     // stdInv = 1 / (v + eps)^0.5
     // N - batch size (product of spatial dimensions)
 

libnd4j/include/ops/declarable/generic/nn/convo/conv1d.cpp

@@ -222,7 +222,7 @@ CUSTOM_OP_IMPL(conv1d_bp, 3, 2, false, 0, 5) {
     auto gradWReshaped   = gradW  ->reshape(gradW->ordering(),  {1, weights->sizeAt(0), weights->sizeAt(1), weights->sizeAt(2)});    // [kW, iC, oC] -> [1, kW, iC, oC]
 
     nd4j::ops::conv2d_bp conv2dBP;
-    const Nd4jStatus status = conv2dBP.execute({&inputReshaped, &weightsReshaped, bias, &gradOReshaped}, {&gradIReshaped, &gradWReshaped, gradB}, {}, {1,kW,  1,sW,  0,pW,  1,dW,  paddingMode,  !isNCW}, {});
+    auto status = conv2dBP.execute({&inputReshaped, &weightsReshaped, bias, &gradOReshaped}, {&gradIReshaped, &gradWReshaped, gradB}, {}, {1,kW,  1,sW,  0,pW,  1,dW,  paddingMode,  !isNCW}, {});
     if (status != ND4J_STATUS_OK)
         return status;
 

libnd4j/include/ops/declarable/generic/nn/dot_product_attention.cpp

@@ -91,7 +91,7 @@ namespace ops  {
         }
 
         nd4j::ops::softmax softmax;
-        softmax.execute({weights}, {weights}, {}, {-2}, {}, true);
+        softmax.execute({weights}, std::vector<NDArray*>{weights}, {}, {-2}, {}, {}, true);
 
         mmul.execute({values, weights}, {output}, {}, {}, {});
 
@@ -189,7 +189,7 @@ namespace ops  {
 
         nd4j::ops::matmul_bp mmul_bp;
         NDArray dLdw(weights.getShapeInfo(), block.workspace());
-        mmul_bp.execute({values, &weights, eps}, {dLdv, &dLdw}, {}, {}, {});
+        mmul_bp.execute({values, &weights, eps}, std::vector<NDArray*>{dLdv, &dLdw}, {}, {}, {});
 
         NDArray dLds(preSoftmax.shapeInfo(), block.workspace());
         nd4j::ops::softmax_bp softmax_bp;
@@ -198,7 +198,7 @@ namespace ops  {
         if(normalization)
             dLds /= factor;
 
-        mmul_bp.execute({keys, queries, &dLds}, {dLdk, dLdq}, {}, {1}, {});
+        mmul_bp.execute({keys, queries, &dLds}, std::vector<NDArray*>{dLdk, dLdq}, {}, {1}, {});
 
         return Status::OK();
     }

libnd4j/include/ops/declarable/generic/nn/multi_head_dot_product_attention.cpp

@@ -239,7 +239,7 @@ namespace ops  {
         auto epsPostReshape = epsPerm.reshape(eps->ordering(), {miniBatchSize * queryCount, outSize});
         nd4j::ops::matmul_bp matmulBp;
         NDArray dLdPreWo(attnResults.shapeInfo(), false, block.launchContext());
-        matmulBp.execute({&attnResults, Wo, &epsPostReshape}, {&dLdPreWo, dLdWo}, {}, {}, {});
+        matmulBp.execute({&attnResults, Wo, &epsPostReshape}, std::vector<NDArray*>{&dLdPreWo, dLdWo}, {}, {}, {});
 
         // dLdAttn
         dLdPreWo.reshapei({miniBatchSize, queryCount, numHeads, projectedValues.sizeAt(2)});

libnd4j/include/ops/declarable/generic/nn/pooling/avgpool2d.cpp

@@ -31,31 +31,28 @@ namespace ops  {
 CUSTOM_OP_IMPL(avgpool2d, 1, 1, false, 0, 10) {
 
     auto input = INPUT_VARIABLE(0);
-
+    auto output = OUTPUT_VARIABLE(0);
-    REQUIRE_TRUE(input->rankOf() == 4, 0, "Input should have rank of 4, but got %i instead", input->rankOf());
 
     // 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 - same mode;
-    auto argI = *(block.getIArguments());
-    auto output = OUTPUT_VARIABLE(0);
 
     const auto kH = INT_ARG(0);
     const auto kW = INT_ARG(1);
     const auto sH = INT_ARG(2);
     const auto sW = INT_ARG(3);
-          int pH = INT_ARG(4);
+          auto pH = INT_ARG(4);
-          int pW = INT_ARG(5);
+          auto pW = INT_ARG(5);
     const auto dH = INT_ARG(6);
     const auto dW = INT_ARG(7);
     const auto isSameMode = static_cast<bool>(INT_ARG(8));
     const auto extraParam0 = INT_ARG(9);
+    const int isNCHW  = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1;       // INT_ARG(10): 0-NCHW, 1-NHWC
 
+    REQUIRE_TRUE(input->rankOf() == 4, 0, "AVGPOOL2D op: input should have rank of 4, but got %i instead", input->rankOf());
     REQUIRE_TRUE(dH != 0 && dW != 0, 0, "AVGPOOL2D op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
 
     int oH = 0;
     int oW = 0;
 
-    int isNCHW  = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1;       // INT_ARG(10): 0-NCHW, 1-NHWC
-
     const int iH = static_cast<int>(isNCHW ? input->sizeAt(2) : input->sizeAt(1));
     const int iW = static_cast<int>(isNCHW ? input->sizeAt(3) : input->sizeAt(2));
 
@@ -207,7 +204,6 @@ CUSTOM_OP_IMPL(avgpool2d_bp, 2, 1, false, 0, 10) {
     }
 
     return Status::OK();
-
 }
 
 DECLARE_SHAPE_FN(avgpool2d_bp) {

libnd4j/include/ops/declarable/generic/nn/pooling/avgpool3d.cpp

@@ -51,14 +51,14 @@ CUSTOM_OP_IMPL(avgpool3dnew, 1, 1, false, 0, 14) {
     int isNCDHW  = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1;       // 0-NCDHW, 1-NDHWC
 
     REQUIRE_TRUE(input->rankOf() == 5, 0, "AVGPOOL3DNEW OP: rank of input array must be equal to 5, but got %i instead !", input->rankOf());
-    REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0, "AVGPOOL3DNEW op: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
+    REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0, "AVGPOOL3DNEW OP: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
 
     int bS, iC, iD, iH, iW, oC, oD, oH, oW;                     // batch size, input channels, input depth/height/width, output channels, output depth/height/width;
     int indIOioC, indIOioD, indWoC, indWiC, indWkD;             // corresponding indexes
     ConvolutionUtils::getSizesAndIndexesConv3d(isNCDHW, *input, *output, bS, iC, iD, iH, iW, oC, oD, oH, oW, indIOioC, indIOioD, indWiC, indWoC, indWkD);
 
     std::string expectedOutputShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,iC,oD,oH,oW,  0,indIOioC,indIOioD,indIOioD+1,indIOioD+2}));
-    REQUIRE_TRUE(expectedOutputShape == ShapeUtils::shapeAsString(output), 0, "AVGPOOL3D op: wrong shape of output array, expected is %s, but got %s instead !", expectedOutputShape.c_str(), ShapeUtils::shapeAsString(output).c_str());
+    REQUIRE_TRUE(expectedOutputShape == ShapeUtils::shapeAsString(output), 0, "AVGPOOL3DNEW OP: wrong shape of output array, expected is %s, but got %s instead !", expectedOutputShape.c_str(), ShapeUtils::shapeAsString(output).c_str());
 
     if(!isNCDHW) {
         input  = new NDArray(input->permute({0, 4, 1, 2, 3}));              // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
@@ -176,8 +176,8 @@ CUSTOM_OP_IMPL(avgpool3dnew_bp, 2, 1, false, 0, 14) {
 
     std::string expectedGradOShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,iC,oD,oH,oW,  0,indIOioC,indIOioD,indIOioD+1,indIOioD+2}));
     std::string expectedGradIShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,iC,iD,iH,iW,  0,indIOioC,indIOioD,indIOioD+1,indIOioD+2}));
-    REQUIRE_TRUE(expectedGradOShape == ShapeUtils::shapeAsString(gradO), 0, "AVGPOOL3D_BP op: wrong shape of output's gradients array (next epsilon), expected is %s, but got %s instead !", expectedGradOShape.c_str(), ShapeUtils::shapeAsString(gradO).c_str());
+    REQUIRE_TRUE(expectedGradOShape == ShapeUtils::shapeAsString(gradO), 0, "AVGPOOL3DNEW_BP op: wrong shape of output's gradients array (next epsilon), expected is %s, but got %s instead !", expectedGradOShape.c_str(), ShapeUtils::shapeAsString(gradO).c_str());
-    REQUIRE_TRUE(expectedGradIShape == ShapeUtils::shapeAsString(gradI), 0, "AVGPOOL3D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !", expectedGradIShape.c_str(), ShapeUtils::shapeAsString(gradI).c_str());
+    REQUIRE_TRUE(expectedGradIShape == ShapeUtils::shapeAsString(gradI), 0, "AVGPOOL3DNEW_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !", expectedGradIShape.c_str(), ShapeUtils::shapeAsString(gradI).c_str());
 
     if(!isNCDHW) {
         input = new NDArray(input->permute({0, 4, 1, 2, 3}));                                   // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]

libnd4j/include/ops/declarable/generic/nn/pooling/maxpool2d.cpp

@@ -32,6 +32,7 @@ namespace ops  {
 //////////////////////////////////////////////////////////////////////////
 // maxpool2d corresponds to poolingMode=0
 CUSTOM_OP_IMPL(maxpool2d, 1, 1, false, 0, 9) {
+
     auto input = INPUT_VARIABLE(0);
 
     REQUIRE_TRUE(input->rankOf() == 4, 0, "MAXPOOL2D OP: input array should have rank of 4, but got %i instead", input->rankOf());

libnd4j/include/ops/declarable/generic/nn/relu_layer.cpp

@@ -40,7 +40,7 @@ namespace nd4j {
             //nd4j_printf("Matrix x(%ix%i), Matrix w(%ix%i), b(1x%i)\n", x->sizeAt(0), x->sizeAt(1), w->sizeAt(0), w->sizeAt(1), b->lengthOf());
 
             nd4j::ops::xw_plus_b op;
-            std::unique_ptr<ResultSet> result(op.execute({x, w, b}, {}, {}, {}));
+            std::unique_ptr<ResultSet> result(op.evaluate({x, w, b}));
             REQUIRE_TRUE(Status::OK() == result->status(), 0, "relu_layer: xw_plus_b op failed on input data.");
 
             auto scalar = block.numT() > 0 ? block.getTArguments()->at(0) : 0.0;

libnd4j/include/ops/declarable/generic/parity_ops/compare_and_bitpack.cpp

@@ -34,7 +34,7 @@ namespace nd4j {
             
             auto tZ = BroadcastHelper::broadcastApply(BROADCAST_BOOL(GreaterThan), x, y, &z0);
             bitcast res;
-            auto status = res.execute({tZ}, {z}, {}, {DataType::UINT8}, {}, false);
+            auto status = res.execute({tZ}, {z}, {}, {DataType::UINT8}, {}, {}, false);
             if (tZ != &z0) {
                 delete tZ;
             }

libnd4j/include/ops/declarable/generic/parity_ops/dynamic_parititon.cpp

@@ -112,7 +112,7 @@ namespace ops {
         NDArray originalIndices(*indices); //->ordering(), indices->shapeInfo(), indices->dataType());
         originalIndices.linspace(0);
         ops::dynamic_partition op;
-        auto res = op.execute({&originalIndices, indices}, {}, {numPartition});
+        auto res = op.evaluate({&originalIndices, indices}, {numPartition});
         REQUIRE_TRUE(res->status() == ND4J_STATUS_OK, 0, "dynamic_partition_bp: Error with dynamic partitioning.");
         ops::dynamic_stitch stichOp;
         std::vector<NDArray*> partitions(numPartition * 2);
@@ -121,7 +121,7 @@ namespace ops {
             partitions[i + numPartition] = gradOutList[i];
         }
 
-        auto result = stichOp.execute(partitions, {}, {numPartition}, {}, false);
+        auto result = stichOp.evaluate(partitions, {numPartition});
         REQUIRE_TRUE(result->status() == ND4J_STATUS_OK, 0, "dynamic_partition_bp: Error with dynamic partitioning.");
         result->at(0)->reshapei(outputList[0]->getShapeAsVector());
         outputList[1]->assign(indices);

libnd4j/include/ops/declarable/generic/parity_ops/embedding_lookup.cpp

@@ -66,7 +66,7 @@ CUSTOM_OP_IMPL(embedding_lookup, 2, 1, false, 0, 1) {
 
         nd4j::ops::gather op;
 
-        std::unique_ptr<ResultSet> result(op.execute({input, indeces}, {}, {0}, {}));
+        std::unique_ptr<ResultSet> result(op.evaluate({input, indeces}, {0}));
         REQUIRE_TRUE(result->status() == Status::OK(), 0, "embedding_lookup: cannot retrieve results from gather op.");
         REQUIRE_TRUE(result->at(0)->isSameShape(output), 0, "embedding_lookup: wrong shape of return from gather op.");
         output->assign(result->at(0));
@@ -94,7 +94,7 @@ DECLARE_SHAPE_FN(embedding_lookup) {
         for (int e = 1; e < outRank; e++)
             shapeInfo[e] = shape::sizeAt(inShapeInfo, e);
 
-        auto outShapeInfo = ConstantShapeHelper::getInstance()->createShapeInfo(block.dataType(), shape::order(inShapeInfo), shapeInfo);
+        auto outShapeInfo = ConstantShapeHelper::getInstance()->createShapeInfo(ArrayOptions::dataType(inShapeInfo), shape::order(inShapeInfo), shapeInfo);
         return SHAPELIST(outShapeInfo);
     }
 

libnd4j/include/ops/declarable/generic/parity_ops/lin_space.cpp

@@ -32,6 +32,11 @@ namespace ops {
         auto finish = INPUT_VARIABLE(1);
         auto numOfElements = INPUT_VARIABLE(2);
 
+        if (numOfElements->e<Nd4jLong>(0) == 1) {
+            output->assign(start);
+            return Status::OK();
+        }
+    
         output->linspace(start->e<double>(0), (finish->e<double>(0) - start->e<double>(0)) / (numOfElements->e<Nd4jLong>(0) - 1.));
         return Status::OK();
     }

libnd4j/include/ops/declarable/generic/parity_ops/onehot.cpp

@@ -74,6 +74,8 @@ namespace nd4j {
         DECLARE_SHAPE_FN(onehot) {
             auto inShape = inputShape->at(0);
 
+            nd4j::DataType dtype = block.numD() > 0 ? D_ARG(0) : nd4j::DataType::FLOAT32;
+
             int depth = -1;
             Nd4jLong axis = -1;
 
@@ -99,7 +101,7 @@ namespace nd4j {
                 shape.push_back(shape::shapeOf(inShape)[e]);
 
             shape.insert(shape.begin() + axis, depth);
-            newShape = ConstantShapeHelper::getInstance()->createShapeInfo(block.dataType(), 'c', rank + 1, shape.data());
+            newShape = ConstantShapeHelper::getInstance()->createShapeInfo(dtype, 'c', rank + 1, shape.data());
 
             return SHAPELIST(newShape);
         }

libnd4j/include/ops/declarable/generic/parity_ops/ones_as.cpp

@@ -25,7 +25,7 @@
 
 namespace nd4j {
     namespace ops {
-        OP_IMPL(ones_as, 1, 1, false) {
+        CUSTOM_OP_IMPL(ones_as, 1, 1, false, 0, 0) {
             auto output = OUTPUT_VARIABLE(0);
 
             output->assign(1);
@@ -33,11 +33,21 @@ namespace nd4j {
             return Status::OK();
         }
 
+        DECLARE_SHAPE_FN(ones_as) {
+            auto in = inputShape->at(0);
+            auto dtype = block.numD() ? D_ARG(0) : ArrayOptions::dataType(in);
+            auto shape = nd4j::ConstantShapeHelper::getInstance()->createShapeInfo(dtype, in);
+
+            nd4j_printf("numD: %i; dtype: %s\n", block.numD(), DataTypeUtils::asString(dtype).c_str());
+
+            return SHAPELIST(shape);
+        }
+
         DECLARE_TYPES(ones_as) {
             getOpDescriptor()
                     ->setAllowedInputTypes(nd4j::DataType::ANY)
                     ->setAllowedOutputTypes(nd4j::DataType::ANY)
-                    ->setSameMode(true);
+                    ->setSameMode(false);
         }
     }
 }

libnd4j/include/ops/declarable/generic/parity_ops/range.cpp

@@ -130,7 +130,7 @@ DECLARE_SHAPE_FN(range) {
     const int numIArgs  = block.getIArguments()->size();    
 
     Nd4jLong steps = 0;
-    nd4j::DataType dataType = nd4j::DataType::INHERIT;
+    nd4j::DataType dataType = block.numD() ? D_ARG(0) : nd4j::DataType::INHERIT;
 
     if (numInArrs > 0) {
         auto isR = INPUT_VARIABLE(0)->isR();
@@ -159,7 +159,9 @@ DECLARE_SHAPE_FN(range) {
             REQUIRE_TRUE(delta != 0, 0, "CUSTOM RANGE OP: delta should not be equal to zero !");
 
             steps = static_cast<Nd4jLong >((limit - start) / delta);
-            dataType = INPUT_VARIABLE(0)->dataType();
+
+            if (!block.numD())
+                dataType = INPUT_VARIABLE(0)->dataType();
 
             if(math::nd4j_abs<double>(start + steps * delta) < math::nd4j_abs<double >(limit))
                 ++steps;
@@ -187,7 +189,9 @@ DECLARE_SHAPE_FN(range) {
             REQUIRE_TRUE(delta != 0, 0, "CUSTOM RANGE OP: delta should not be equal to zero !");
 
             steps = static_cast<Nd4jLong >((limit - start) / delta);
-            dataType = INPUT_VARIABLE(0)->dataType();
+
+            if (!block.numD())
+                dataType = INPUT_VARIABLE(0)->dataType();
 
             if(math::nd4j_abs<double>(start + steps * delta) < math::nd4j_abs<double >(limit))
                 ++steps;
@@ -214,10 +218,12 @@ DECLARE_SHAPE_FN(range) {
 
         REQUIRE_TRUE(delta != 0, 0, "CUSTOM RANGE OP: delta should not be equal to zero !");
 
-        if (limit > DataTypeUtils::max<int>())
+        if (!block.numD()) {
-            dataType = nd4j::DataType::INT64;
+            if (limit > DataTypeUtils::max<int>())
-        else
+                dataType = nd4j::DataType::INT64;
-            dataType = nd4j::DataType::INT32;
+            else
+                dataType = nd4j::DataType::INT32;
+        }
 
         steps = (limit - start) / delta;
 
@@ -248,10 +254,13 @@ DECLARE_SHAPE_FN(range) {
         REQUIRE_TRUE(delta != 0, 0, "CUSTOM RANGE OP: delta should not be equal to zero !");
 
         steps = static_cast<Nd4jLong >((limit - start) / delta);
-        if (Environment::getInstance()->precisionBoostAllowed())
+
-            dataType = nd4j::DataType::DOUBLE;
+        if (!block.numD()) {
-        else
+            if (Environment::getInstance()->precisionBoostAllowed())
-            dataType = Environment::getInstance()->defaultFloatDataType();
+                dataType = nd4j::DataType::DOUBLE;
+            else
+                dataType = Environment::getInstance()->defaultFloatDataType();
+        }
 
         if(math::nd4j_abs<double>(start + steps * delta) < math::nd4j_abs<double >(limit))
             ++steps;

libnd4j/include/ops/declarable/generic/parity_ops/solve.cpp

@@ -0,0 +1,75 @@
+/*******************************************************************************
+ * Copyright (c) 2020 Konduit, K.K.
+ *
+ * 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
+ ******************************************************************************/
+
+//
+// Created by GS <sgazeos@gmail.com> at 01/22/2020
+//
+
+#include <op_boilerplate.h>
+#if NOT_EXCLUDED(OP_solve)
+
+#include <ops/declarable/CustomOperations.h>
+#include <ops/declarable/helpers/solve.h>
+namespace nd4j {
+    namespace ops {
+        CUSTOM_OP_IMPL(solve, 2, 1, false, 0, 0) {
+            auto a = INPUT_VARIABLE(0);
+            auto b = INPUT_VARIABLE(1);
+            auto z = OUTPUT_VARIABLE(0);
+
+            bool useAdjoint = false;
+
+            if (block.numB() > 0) {
+                useAdjoint = B_ARG(0);
+            }
+
+            REQUIRE_TRUE(a->rankOf() >=2, 0, "solve: The rank of input left tensor should not be less than 2, but %i is given", a->rankOf());
+            REQUIRE_TRUE(b->rankOf() >=2, 0, "solve: The rank of input right tensor should not be less than 2, but %i is given", b->rankOf());
+
+            REQUIRE_TRUE(a->sizeAt(-1) == a->sizeAt(-2), 0, "solve: The last two dimmensions should be equal, but %i and %i are given", a->sizeAt(-1), a->sizeAt(-2));
+            REQUIRE_TRUE(a->sizeAt(-1) == b->sizeAt(-2), 0, "solve: The last dimmension of left part should be equal to prelast of right part, but %i and %i are given", a->sizeAt(-1), b->sizeAt(-2));
+            auto input = a;
+            if (useAdjoint) {
+                auto adjointA = a->ulike();
+                helpers::adjointMatrix(block.launchContext(), a, &adjointA);
+                input = new NDArray(adjointA); //.detach();
+            };
+
+            auto res = helpers::solveFunctor(block.launchContext(), input, b, useAdjoint, z);
+            if (input != a)
+                delete input;
+
+            return Status::OK();
+        }
+        
+        DECLARE_SHAPE_FN(solve) {
+            auto in0 = inputShape->at(1);
+            auto in1 = inputShape->at(1);
+            auto luShape = ShapeBuilders::copyShapeInfoAndType(in1, in0, true, block.workspace());
+
+            return SHAPELIST(CONSTANT(luShape));
+        }
+
+        DECLARE_TYPES(solve) {
+            getOpDescriptor()
+                    ->setAllowedInputTypes({ALL_FLOATS})
+                    ->setAllowedOutputTypes({ALL_FLOATS})
+                    ->setSameMode(false);
+        }
+    }
+}
+
+#endif

libnd4j/include/ops/declarable/generic/parity_ops/zeros_as.cpp

@@ -25,7 +25,7 @@
 
 namespace nd4j {
     namespace ops {
-        OP_IMPL(zeros_as, 1, 1, false) {
+        CUSTOM_OP_IMPL(zeros_as, 1, 1, false, 0, 0) {
             auto out = OUTPUT_VARIABLE(0);
 
             out->assign(0); // output is filled by zero by default
@@ -35,11 +35,20 @@ namespace nd4j {
         DECLARE_SYN(zeroslike, zeros_as);
         DECLARE_SYN(zeros_like, zeros_as);
 
+
+        DECLARE_SHAPE_FN(zeros_as) {
+            auto in = inputShape->at(0);
+            auto dtype = block.numD() ? D_ARG(0) : ArrayOptions::dataType(in);
+            auto shape = nd4j::ConstantShapeHelper::getInstance()->createShapeInfo(dtype, in);
+
+            return SHAPELIST(shape);
+        }
+
         DECLARE_TYPES(zeros_as) {
             getOpDescriptor()
                     ->setAllowedInputTypes(nd4j::DataType::ANY)
                     ->setAllowedOutputTypes(nd4j::DataType::ANY)
-                    ->setSameMode(true);
+                    ->setSameMode(false);
         }
     }
 }

libnd4j/include/ops/declarable/generic/recurrent/dynamicBidirectionalRNN.cpp

@@ -84,7 +84,7 @@ CUSTOM_OP_IMPL(dynamic_bidirectional_rnn, 7, 4, false, 0, 0) {
 
     // forward steps
     nd4j::ops::dynamic_rnn dynamicRnn;
-    auto resultsFW = dynamicRnn.execute({x, WxFW, WhFW, bFW, h0FW, maxTimeStep}, {}, {timeMajor}, {}, false, x->dataType());
+    auto resultsFW = dynamicRnn.evaluate({x, WxFW, WhFW, bFW, h0FW, maxTimeStep}, {timeMajor});
     hFW->assign(resultsFW->at(0));                              // [time x bS x numUnitsFW] or [bS x time x numUnitsFW]
     hFWFinal->assign(resultsFW->at(1));
 
@@ -97,17 +97,17 @@ CUSTOM_OP_IMPL(dynamic_bidirectional_rnn, 7, 4, false, 0, 0) {
 
     // reverse x     
     nd4j::ops::reverse_sequence reverse;
-    auto resultsIn = timeMajor ? reverse.execute({x, seqLen}, {}, {0, 1}, {}, false, x->dataType()) : reverse.execute({x, seqLen}, {}, {1, 0}, {}, false, x->dataType());
+    auto resultsIn = timeMajor ? reverse.evaluate({x, seqLen}, {0, 1}) : reverse.evaluate({x, seqLen}, {1, 0});
     REQUIRE_TRUE (resultsIn->status() == ND4J_STATUS_OK, 0, "dynamic_bidirectional_rnn: there is a problem with reverse on the sequence.");
     auto revInput = resultsIn->at(0);
 
     // backward steps    
-    auto resultsBW = dynamicRnn.execute({revInput, WxBW, WhBW, bBW, h0BW, maxTimeStep}, {}, {timeMajor}, {});
+    auto resultsBW = dynamicRnn.evaluate({revInput, WxBW, WhBW, bBW, h0BW, maxTimeStep}, {timeMajor});
     auto hBWtemp = resultsBW->at(0);					           // [time x bS x numUnitsBW] or [ bS x time xnumUnitsBW]
     hBWFinal->assign(resultsBW->at(1));
 
     // reverse hBWtemp 
-    auto resultsOut = timeMajor ? reverse.execute({hBWtemp, seqLen}, {}, {0, 1}, {}) : reverse.execute({hBWtemp, seqLen}, {}, {1, 0}, {});
+    auto resultsOut = timeMajor ? reverse.evaluate({hBWtemp, seqLen}, {0, 1}) : reverse.evaluate({hBWtemp, seqLen}, {1, 0});
     hBW->assign(resultsOut->at(0));
     
 	delete resultsOut;

libnd4j/include/ops/declarable/generic/shape/tile_to_shape.cpp

@@ -48,7 +48,7 @@ namespace ops {
 
         auto conv = ArrayUtils::toLongVector(*block.getIArguments());
 
-        auto newShape = ConstantShapeHelper::getInstance()->createShapeInfo(block.dataType(), shape::order(in), conv);
+        auto newShape = ConstantShapeHelper::getInstance()->createShapeInfo(ArrayOptions::dataType(in), shape::order(in), conv);
 
         return SHAPELIST(newShape);
     }

libnd4j/include/ops/declarable/headers/parity_ops.h

@@ -487,7 +487,7 @@ namespace nd4j {
          *
          */
         #if NOT_EXCLUDED(OP_zeros_as)
-        DECLARE_OP(zeros_as, 1, 1, false);
+        DECLARE_CUSTOM_OP(zeros_as, 1, 1, false, 0, 0);
         #endif
 
         /**
@@ -497,7 +497,7 @@ namespace nd4j {
          *
          */
         #if NOT_EXCLUDED(OP_ones_as)
-        DECLARE_OP(ones_as, 1, 1, false);
+        DECLARE_CUSTOM_OP(ones_as, 1, 1, false, 0, 0);
         #endif
 
         /**
@@ -1076,6 +1076,24 @@ namespace nd4j {
         DECLARE_CUSTOM_OP(triangular_solve, 2, 1, true, 0, 0);
         #endif
 
+        /**
+         * solve op. - solve systems of linear equations - general method.
+         *
+         * input params:
+         *    0 - the tensor with dimension (x * y * z * ::: * M * M) - left parts of equations
+         *    1 - the tensor with dimension (x * y * z * ::: * M * K) - right parts of equations
+         *
+         * boolean args:
+         *    0 - adjoint - default is false (optional) - indicate input matrix or its adjoint (hermitian addition) should be used
+         *
+         * return value:
+         *    tensor with dimension (x * y * z * ::: * M * K) with solutions
+         *
+         */
+        #if NOT_EXCLUDED(OP_solve)
+        DECLARE_CUSTOM_OP(solve, 2, 1, true, 0, 0);
+        #endif
+
         /**
          * lu op. - make LUP decomposition of given batch of 2D square matricies
          *

libnd4j/include/ops/declarable/helpers/cpu/lup.cpp

@@ -237,25 +237,65 @@ namespace helpers {
         samediff::Threads::parallel_tad(loop, currentRow + 1, rowNum, 1);
     }
 
+    template <typename T>
+    static void doolitleLU(LaunchContext* context, NDArray* compound, Nd4jLong rowNum) {
+        auto input = compound->dup();
+        compound->nullify();
+
+        // Decomposing matrix into Upper and Lower
+        // triangular matrix
+        for (auto i = 0; i < rowNum; i++) {
+
+            // Upper Triangular
+            for (auto k = i; k < rowNum; k++) {
+
+                // Summation of L(i, j) * U(j, k)
+                int sum = 0;
+                for (int j = 0; j < i; j++)
+                    sum += compound->t<T>(i,j) * compound->t<T>(j,k);
+
+                // Evaluating U(i, k)
+                compound->t<T>(i, k) = input.t<T>(i, k) - sum;
+            }
+
+            // Lower Triangular
+            for (int k = i + 1; k < rowNum; k++) {
+                // Summation of L(k, j) * U(j, i)
+                int sum = 0;
+                for (int j = 0; j < i; j++)
+                    sum += compound->t<T>(k,j) * compound->t<T>(j, i);
+
+                // Evaluating L(k, i)
+                compound->t<T>(k, i) = (input.t<T>(k, i) - sum) / compound->t<T>(i,i);
+            }
+        }
+    }
+
     template <typename T, typename I>
     static void luNN_(LaunchContext *context, NDArray* compound, NDArray* permutation, Nd4jLong rowNum) {
 
         //const int rowNum = compound->rows();
 //        const int columnNum = output->columns();
-        permutation->linspace(0);
+        if (permutation) { // LUP algorithm
-        auto permutationBuf = permutation->bufferAsT<I>(); //dataBuffer()->primaryAsT<I>();
+            permutation->linspace(0);
-        auto compoundBuf = compound->bufferAsT<T>();
+            auto permutationBuf = permutation->bufferAsT<I>(); //dataBuffer()->primaryAsT<I>();
-        auto compoundShape = compound->shapeInfo();
+            auto compoundBuf = compound->bufferAsT<T>();
-        auto permutationShape = permutation->shapeInfo();
+            auto compoundShape = compound->shapeInfo();
-        for (auto i = 0; i < rowNum - 1; i++) {
+            auto permutationShape = permutation->shapeInfo();
-            auto pivotIndex = argmaxCol(i, compoundBuf, compoundShape);
+            for (auto i = 0; i < rowNum - 1; i++) {
-            if (pivotIndex < 0) {
+                auto pivotIndex = argmaxCol(i, compoundBuf, compoundShape);
-                throw std::runtime_error("helpers::luNN_: input matrix is singular.");
+                if (pivotIndex < 0) {
-            }
+                    throw std::runtime_error("helpers::luNN_: input matrix is singular.");
-            math::nd4j_swap(permutationBuf[shape::getIndexOffset(i, permutationShape)], permutationBuf[shape::getIndexOffset(pivotIndex, permutationShape)]);
+                }
-            swapRows(compoundBuf, compoundShape, i, pivotIndex);
+                math::nd4j_swap(permutationBuf[shape::getIndexOffset(i, permutationShape)],
+                                permutationBuf[shape::getIndexOffset(pivotIndex, permutationShape)]);
+                swapRows(compoundBuf, compoundShape, i, pivotIndex);
 
-            processColumns(i, rowNum, compoundBuf, compoundShape);
+                processColumns(i, rowNum, compoundBuf, compoundShape);
+            }
+        }
+        else { // Doolitle algorithm with LU decomposition
+            doolitleLU<T>(context, compound, rowNum);
         }
     }
 
@@ -265,17 +305,20 @@ namespace helpers {
 
         output->assign(input); // fill up output tensor with zeros
         ResultSet outputs = output->allTensorsAlongDimension({-2, -1});
-        ResultSet permutations = permutationVectors->allTensorsAlongDimension({-1});
+        ResultSet permutations;
+        if (permutationVectors)
+            permutations = permutationVectors->allTensorsAlongDimension({-1});
+
         auto loop = PRAGMA_THREADS_FOR {
             for (auto i = start; i < stop; i += increment) {
-                luNN_<T, I>(context, outputs.at(i), permutations.at(i), n);
+                luNN_<T, I>(context, outputs.at(i), permutationVectors?permutations.at(i):nullptr, n);
             }
         };
         samediff::Threads::parallel_for(loop, 0, outputs.size(), 1);
     }
 
     void lu(LaunchContext *context, NDArray* input, NDArray* output, NDArray* permutation) {
-        BUILD_DOUBLE_SELECTOR(input->dataType(), permutation->dataType(), lu_, (context, input, output, permutation), FLOAT_TYPES, INDEXING_TYPES);
+        BUILD_DOUBLE_SELECTOR(input->dataType(), permutation?permutation->dataType():DataType::INT32, lu_, (context, input, output, permutation), FLOAT_TYPES, INDEXING_TYPES);
     }
 
 //    BUILD_DOUBLE_TEMPLATE(template NDArray lu_, (LaunchContext *context, NDArray* input, NDArray* output, NDArray* permutation), FLOAT_TYPES, INDEXING_TYPES);

libnd4j/include/ops/declarable/helpers/cpu/solve.cpp

@@ -0,0 +1,100 @@
+/*******************************************************************************
+ * Copyright (c) 2020 Konduit, K.K.
+ *
+ * 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 GS <sgazeos@gmail.com>
+//
+#include <op_boilerplate.h>
+#include <NDArray.h>
+#include <NDArrayFactory.h>
+#include <execution/Threads.h>
+#include <helpers/MmulHelper.h>
+
+#include "../triangular_solve.h"
+#include "../lup.h"
+#include "../solve.h"
+
+namespace nd4j {
+namespace ops {
+namespace helpers {
+
+// --------------------------------------------------------------------------------------------------------------------------------------- //
+    template <typename T>
+    static void adjointMatrix_(nd4j::LaunchContext* context, NDArray const* input, NDArray* output) {
+        auto inputPart = input->allTensorsAlongDimension({-2, -1});
+        auto outputPart = output->allTensorsAlongDimension({-2, -1});
+        output->assign(input);
+        auto batchLoop = PRAGMA_THREADS_FOR {
+            for (auto batch = start; batch < stop; batch += increment) {
+                for (auto r = 0; r < input->rows(); r++) {
+                    for (auto c = 0; c < r; c++) {
+                        math::nd4j_swap(outputPart[batch]->t<T>(r, c) , outputPart[batch]->t<T>(c, r));
+                    }
+                }
+            }
+        };
+        samediff::Threads::parallel_tad(batchLoop, 0, inputPart.size(), 1);
+    }
+
+// --------------------------------------------------------------------------------------------------------------------------------------- //
+    template <typename T>
+    static int solveFunctor_(nd4j::LaunchContext * context, NDArray* leftInput, NDArray* rightInput, bool const adjoint, NDArray* output) {
+
+        // stage 1: LU decomposition batched
+        auto leftOutput = leftInput->ulike();
+        auto permuShape = rightInput->getShapeAsVector(); permuShape.pop_back();
+        auto permutations = NDArrayFactory::create<int>('c', permuShape, context);
+        helpers::lu(context, leftInput, &leftOutput, &permutations);
+        auto P = leftInput->ulike(); //permutations batched matrix
+        P.nullify(); // to fill up matricies with zeros
+        auto PPart = P.allTensorsAlongDimension({-2,-1});
+        auto permutationsPart = permutations.allTensorsAlongDimension({-1});
+
+        for (auto batch = 0; batch < permutationsPart.size(); ++batch) {
+            for (auto row = 0; row < PPart[batch]->rows(); ++row) {
+                PPart[batch]->t<T>(row, permutationsPart[batch]->t<int>(row)) = T(1.f);
+            }
+        }
+
+        auto leftLower = leftOutput.dup();
+        auto rightOutput = rightInput->ulike();
+        auto rightPermuted = rightOutput.ulike();
+        MmulHelper::matmul(&P, rightInput, &rightPermuted, 0, 0);
+        ResultSet leftLowerPart = leftLower.allTensorsAlongDimension({-2, -1});
+        for (auto i = 0; i < leftLowerPart.size(); i++) {
+            for (auto r = 0; r < leftLowerPart[i]->rows(); r++)
+                leftLowerPart[i]->t<T>(r,r) = (T)1.f;
+        }
+        // stage 2: triangularSolveFunctor for Lower with given b
+        helpers::triangularSolveFunctor(context, &leftLower, &rightPermuted, true, false, &rightOutput);
+        // stage 3: triangularSolveFunctor for Upper with output of previous stage
+        helpers::triangularSolveFunctor(context, &leftOutput, &rightOutput, false, false, output);
+
+        return Status::OK();
+    }
+
+// --------------------------------------------------------------------------------------------------------------------------------------- //
+    int solveFunctor(nd4j::LaunchContext * context, NDArray* leftInput, NDArray* rightInput, bool const adjoint, NDArray* output) {
+        BUILD_SINGLE_SELECTOR(leftInput->dataType(), return solveFunctor_, (context, leftInput, rightInput, adjoint, output), FLOAT_TYPES);
+    }
+// --------------------------------------------------------------------------------------------------------------------------------------- //
+    void adjointMatrix(nd4j::LaunchContext* context, NDArray const* input, NDArray* output) {
+        BUILD_SINGLE_SELECTOR(input->dataType(), adjointMatrix_, (context, input, output), FLOAT_TYPES);
+    }
+// --------------------------------------------------------------------------------------------------------------------------------------- //
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/triangular_solve.cpp

@@ -41,13 +41,16 @@ namespace helpers {
     template <typename T>
     static void lowerTriangularSolve(nd4j::LaunchContext * context, NDArray* leftInput, NDArray* rightInput, bool adjoint, NDArray* output) {
         auto rows = leftInput->rows();
+        auto cols = rightInput->columns();
         //output->t<T>(0,0) = rightInput->t<T>(0,0) / leftInput->t<T>(0,0);
         for (auto r = 0; r < rows; r++) {
-            auto sum = rightInput->t<T>(r, 0);
+            for (auto j = 0; j < cols; j++) {
-            for (auto c = 0; c < r; c++) {
+                auto sum = rightInput->t<T>(r, j);
-                sum -= leftInput->t<T>(r,c) * output->t<T>(c, 0);
+                for (auto c = 0; c < r; c++) {
+                    sum -= leftInput->t<T>(r, c) * output->t<T>(c, j);
+                }
+                output->t<T>(r, j) = sum / leftInput->t<T>(r, r);
             }
-            output->t<T>(r, 0) = sum / leftInput->t<T>(r, r);
         }
     }
 
@@ -68,13 +71,15 @@ namespace helpers {
     template <typename T>
     static void upperTriangularSolve(nd4j::LaunchContext * context, NDArray* leftInput, NDArray* rightInput, bool adjoint, NDArray* output) {
         auto rows = leftInput->rows();
-
+        auto cols = rightInput->columns();
         for (auto r = rows; r > 0; r--) {
-            auto sum = rightInput->t<T>(r - 1, 0);
+            for (auto j = 0; j < cols; j++) {
-            for (auto c = r; c < rows; c++) {
+                auto sum = rightInput->t<T>(r - 1, j);
-                sum -= leftInput->t<T>(r - 1, c) * output->t<T>(c, 0);
+                for (auto c = r; c < rows; c++) {
+                    sum -= leftInput->t<T>(r - 1, c) * output->t<T>(c, j);
+                }
+                output->t<T>(r - 1, j) = sum / leftInput->t<T>(r - 1, r - 1);
             }
-            output->t<T>(r - 1, 0) = sum / leftInput->t<T>(r - 1, r - 1);
         }
     }