Rng tweaks (#479)

* initial commit

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

* Java Random.getFloat()/getDouble() methods mapped to C++

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

* Refactored relativeT for float and double data types.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored float relativeT method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored relativeT

Signed-off-by: shugeo <sgazeos@gmail.com>

* - additional rng tests
- float/double uniform generation methos slightly changed

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

* use bitset instead of manual conversion

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

* rollback valueBits changes

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

* remove unused shapelist

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

* update KMeans ground truth test

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

* dedicated union to make MSVC happy

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

* minor tweaks

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

* .seh_savexmm workaround?

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

* don't use march=native in tests on windows

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

Co-authored-by: shugeo <sgazeos@gmail.com>

deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kmeans/KMeansTest.java

@@ -273,10 +273,10 @@ public class KMeansTest extends BaseDL4JTest {
             ClusterSet clusterSet = kMeansClustering.applyTo(points);
 
             double[] centroid1 = {2.44e8,    2.71e8,    2.98e8,    3.25e8};
-            double[] centroid2 = {5.14e8,    5.41e8,    5.68e8,    5.95e8};
+            double[] centroid2 = {1000000.0, 2.8E7, 5.5E7, 8.2E7};
-            double[] centroid3 = {1000000.0, 2.8E7, 5.5E7, 8.2E7};
+            double[] centroid3 = {5.95E8,    6.22e8,    6.49e8,    6.76e8};
-            double[] centroid4 = {7.03E8, 7.3E8, 7.57E8, 7.84E8};
+            double[] centroid4 = {3.79E8, 4.06E8, 4.33E8, 4.6E8};
-            double[] centroid5 = {3.79E8, 4.06E8, 4.33E8, 4.6E8};
+            double[] centroid5 = {5.5E7, 8.2E7, 1.09E8, 1.36E8};
 
             assertArrayEquals(centroid1, clusterSet.getClusters().get(0).getCenter().getArray().toDoubleVector(), 1e-4);
             assertArrayEquals(centroid2, clusterSet.getClusters().get(1).getCenter().getArray().toDoubleVector(), 1e-4);

libnd4j/include/array/NDArray.hXX

@@ -1671,11 +1671,11 @@ void NDArray::printLinearBuffer() const {
     }
     else if(this->dataType() == sd::DataType::FLOAT32) {
         for(Nd4jLong e = 0; e < len; e++)
-            printf("%.3f, ", this->bufferAsT<float>()[e * ews]);
+            printf("%.8f, ", this->bufferAsT<float>()[e * ews]);
     }
     else if(this->dataType() == sd::DataType::DOUBLE) {
         for(Nd4jLong e = 0; e < len; e++)
-            printf("%.3f, ", this->bufferAsT<double>()[e * ews]);
+            printf("%.8f, ", this->bufferAsT<double>()[e * ews]);
     }
     else
         throw std::invalid_argument("NDArray::printLinearBuffer: not implemented yet for this data type !");
@@ -1773,7 +1773,7 @@ void NDArray::printIndexedBuffer(const char* msg, Nd4jLong limit) const {
         if (this->isZ())
             printf("%lld\n", this->e<Nd4jLong>(0));
         else if (this->isR())
-            printf("%f\n", this->e<float>(0));
+            printf("%.8f\n", this->e<float>(0));
         else if (this->isB()) {
             printf("%s\n", this->e<bool>(0)?"true":"false");
         }

libnd4j/include/graph/RandomGenerator.h

@@ -22,6 +22,7 @@
 #define LIBND4J_GRAPH_RNG_H
 
 #include <types/u64.h>
+#include <types/u32.h>
 #include <system/pointercast.h>
 #include <system/op_boilerplate.h>
 #include <system/dll.h>
@@ -29,6 +30,7 @@
 #include <array/DataTypeUtils.h>
 #include <helpers/logger.h>
 #include <stdexcept>
+#include <math/templatemath.h>
 
 #ifdef __CUDACC__
 #include <cuda.h>
@@ -79,9 +81,9 @@ namespace sd {
              */
             static FORCEINLINE Nd4jLong currentMilliseconds();
 
-
+        public:
-            FORCEINLINE _CUDA_HD uint32_t xoroshiro32(Nd4jLong index);
+            FORCEINLINE _CUDA_HD uint32_t xoroshiro32(uint64_t index);
-            FORCEINLINE _CUDA_HD uint64_t xoroshiro64(Nd4jLong index);
+            FORCEINLINE _CUDA_HD uint64_t xoroshiro64(uint64_t index);
 
             /**
              * This method returns integer value between 0 and MAX_UINT
@@ -119,7 +121,7 @@ namespace sd {
             FORCEINLINE _CUDA_HD int relativeInt(Nd4jLong index);
             FORCEINLINE _CUDA_HD Nd4jLong relativeLong(Nd4jLong index);
 
-            FORCEINLINE _CUDA_HD void rewindH(Nd4jLong steps);
+            FORCEINLINE _CUDA_HD void rewindH(uint64_t steps);
 
             /**
              * These methods set up only node states, with non-changed root ones
@@ -172,6 +174,24 @@ namespace sd {
             return v;
         }
 
+        template <>
+        _CUDA_HD FORCEINLINE float RandomGenerator::relativeT<float>(Nd4jLong index) {
+            u32 u;
+            u._u32 = (0x3f800000 | (this->xoroshiro32(index) >> 9));
+            return u._f32 - 1.0f;
+        }
+
+        template <>
+        _CUDA_HD FORCEINLINE double RandomGenerator::relativeT<double>(Nd4jLong index) {
+#ifdef __DOUBLE_RNG__
+          u64 u;
+          u._ulong = ((UINT64_C(0x3FF) << 52) | (this->xoroshiro64(index) >> 12));
+          return u._double - 1.0;
+#else
+          return (double) relativeT<float>(index);
+#endif
+        }
+
         template <>
         _CUDA_HD FORCEINLINE uint64_t RandomGenerator::relativeT<uint64_t>(Nd4jLong index) {
             return this->xoroshiro64(index);
@@ -184,16 +204,14 @@ namespace sd {
 
         template <>
         _CUDA_HD FORCEINLINE int RandomGenerator::relativeT<int>(Nd4jLong index) {
-            auto x = this->relativeT<uint32_t>(index);
+            auto r = relativeT<uint32_t>(index);
-            auto r = static_cast<int>(x % DataTypeUtils::max<int>());
+            return r <= DataTypeUtils::max<int>() ? r : r % DataTypeUtils::max<int>();
-            return r;
         }
 
         template <>
         _CUDA_HD FORCEINLINE Nd4jLong RandomGenerator::relativeT<Nd4jLong>(Nd4jLong index) {
-            auto x = this->relativeT<uint64_t>(index);
+            auto r = relativeT<uint64_t>(index);
-            auto r = static_cast<Nd4jLong>(x % DataTypeUtils::max<Nd4jLong>());
+            return r <= DataTypeUtils::max<Nd4jLong>() ? r : r % DataTypeUtils::max<Nd4jLong>();
-            return r;
         }
 
         template <typename T>
@@ -220,24 +238,18 @@ namespace sd {
         template <typename T>
         _CUDA_HD FORCEINLINE T RandomGenerator::relativeT(Nd4jLong index) {
             // This is default implementation for floating point types
-#ifdef __DOUBLE_RNG__            
+            return static_cast<T>(relativeT<float>(index));
-            auto i = static_cast<double>(this->relativeT<uint64_t>(index));
-            auto r = i / static_cast<double>(DataTypeUtils::max<uint64_t>());
-            return static_cast<T>(r);
-#else            
-            auto i = static_cast<float>(this->relativeT<uint32_t>(index));            
-            auto r = i / static_cast<float>(DataTypeUtils::max<uint32_t>());
-            return static_cast<T>(r);
-#endif
         }
 
 
         _CUDA_HD FORCEINLINE int RandomGenerator::relativeInt(Nd4jLong index) {
-            return relativeT<int>(index);
+            auto r = relativeT<uint32_t>(index);
+            return r <= DataTypeUtils::max<int>() ? r : r % DataTypeUtils::max<int>();
         }
 
         _CUDA_HD FORCEINLINE Nd4jLong RandomGenerator::relativeLong(Nd4jLong index) {
-            return relativeT<Nd4jLong>(index);
+            auto r = relativeT<uint64_t>(index);
+            return r <= DataTypeUtils::max<Nd4jLong>() ? r : r % DataTypeUtils::max<Nd4jLong>();
         }
 
         //////
@@ -249,8 +261,12 @@ namespace sd {
             return (x << k) | (x >> (64 - k));
         }
 
-        _CUDA_HD FORCEINLINE uint32_t RandomGenerator::xoroshiro32(Nd4jLong index) {
+        static FORCEINLINE _CUDA_HD uint32_t next(uint32_t s0, uint32_t s1, uint32_t s2, uint32_t s3) {
+          const uint32_t result = rotl(s0 + s3, 7) + s0;
+          return result;
+        }
 
+        _CUDA_HD FORCEINLINE uint32_t RandomGenerator::xoroshiro32(uint64_t index) {
             auto s0 = _rootState._ulong;
             auto s1 = _nodeState._ulong;
 
@@ -265,23 +281,14 @@ namespace sd {
             return rotl(*pHalf * 0x9E3779BB, 5) * 5;
         }
 
-        _CUDA_HD FORCEINLINE uint64_t RandomGenerator::xoroshiro64(Nd4jLong index) {
+        _CUDA_HD FORCEINLINE uint64_t RandomGenerator::xoroshiro64(uint64_t index) {
-            auto s0 = _rootState._ulong;
+            uint64_t upper = ((uint64_t) xoroshiro32(index)) << 32;
-            auto s1 = _nodeState._ulong;
+            uint32_t lower = xoroshiro32(sd::math::nd4j_rotl<uint64_t>(index, 32));
-
+            return upper + lower;
-            // xor by idx
-            s0 |= ((index + 2) * (s1 + 24243287));
-            s1 ^= ((index + 2) * (s0 + 723829));
-
-            // since we're not modifying state - do rotl step right here
-            s1 ^= s0;
-            s0 = rotl(s0, 55) ^ s1 ^ (s1 << 14);
-            s1 = rotl(s1, 36);
-
-            return s0 + s1;
         }
 
-        _CUDA_HD FORCEINLINE void RandomGenerator::rewindH(Nd4jLong steps) {
+        _CUDA_HD FORCEINLINE void RandomGenerator::rewindH(uint64_t steps) {
+          // we only update node state, if any
           auto s0 = _nodeState._du32._v0;
           auto s1 = _nodeState._du32._v1;
 

libnd4j/include/helpers/StringUtils.h

@@ -44,6 +44,14 @@ namespace sd {
             return os.str();
         }
 
+        /**
+         * These methods convert integer values to string with 0s and 1s
+         * @param value
+         * @return
+         */
+        template <typename T>
+        static std::string bitsToString(T value);
+
         /**
          * This method just concatenates error message with a given graphId
          * @param message
@@ -137,6 +145,9 @@ namespace sd {
         * @return boolean status
         */
         static bool u32StringToU8String(const std::u32string& u32, std::string& u8);
+
+        template <typename T>
+        static std::string vectorToString(const std::vector<T> &vec);
     };
 }
 

libnd4j/include/helpers/impl/BitwiseUtils.cpp

@@ -49,7 +49,6 @@ namespace sd {
         return -1;
     }
 
-
     std::vector<int> BitwiseUtils::valueBits(int holder) {
         std::vector<int> bits;
         if (holder == 0) {
@@ -59,7 +58,6 @@ namespace sd {
           return bits;
         }
 
-
 #ifdef REVERSE_BITS
         for (int e = 32; e >= 0; e--) {
 #else

libnd4j/include/helpers/impl/StringUtils.cpp

@@ -21,7 +21,9 @@
 //
 
 #include <helpers/StringUtils.h>
+#include <helpers/BitwiseUtils.h>
 #include <exceptions/datatype_exception.h>
+#include <bitset>
 
 namespace sd {
     static FORCEINLINE bool match(const uint8_t *haystack, const uint8_t *needle, uint64_t length) {
@@ -32,6 +34,17 @@ namespace sd {
         return true;
     }
 
+    template <typename T>
+    std::string StringUtils::bitsToString(T value) {
+      return std::bitset<sizeof(T) * 8>(value).to_string();
+    }
+
+template std::string StringUtils::bitsToString(int value);
+template std::string StringUtils::bitsToString(uint32_t value);
+template std::string StringUtils::bitsToString(Nd4jLong value);
+template std::string StringUtils::bitsToString(uint64_t value);
+
+
     uint64_t StringUtils::countSubarrays(const void *vhaystack, uint64_t haystackLength, const void *vneedle, uint64_t needleLength) {
         auto haystack = reinterpret_cast<const uint8_t*>(vhaystack);
         auto needle = reinterpret_cast<const uint8_t*>(vneedle);
@@ -155,4 +168,17 @@ namespace sd {
         return true;
     }
 
+  template<typename T>
+  std::string StringUtils::vectorToString(const std::vector<T> &vec) {
+    std::string result;
+    for (auto v:vec)
+      result += valueToString<T>(v);
+
+    return result;
+  }
+
+  template std::string StringUtils::vectorToString(const std::vector<int> &vec);
+  template std::string StringUtils::vectorToString(const std::vector<Nd4jLong> &vec);
+  template std::string StringUtils::vectorToString(const std::vector<int16_t> &vec);
+  template std::string StringUtils::vectorToString(const std::vector<uint32_t> &vec);
 }

libnd4j/include/legacy/NativeOps.h

@@ -1606,6 +1606,8 @@ ND4J_EXPORT OpaqueRandomGenerator* createRandomGenerator(Nd4jLong rootSeed = 0,
 ND4J_EXPORT Nd4jLong getRandomGeneratorRootState(OpaqueRandomGenerator* ptr);
 ND4J_EXPORT Nd4jLong getRandomGeneratorNodeState(OpaqueRandomGenerator* ptr);
 ND4J_EXPORT void setRandomGeneratorStates(OpaqueRandomGenerator* ptr, Nd4jLong rootSeed = 0, Nd4jLong nodeSeed = 0);
+ND4J_EXPORT float getRandomGeneratorRelativeFloat(OpaqueRandomGenerator* ptr, Nd4jLong index);
+ND4J_EXPORT double getRandomGeneratorRelativeDouble(OpaqueRandomGenerator* ptr, Nd4jLong index);
 ND4J_EXPORT int getRandomGeneratorRelativeInt(OpaqueRandomGenerator* ptr, Nd4jLong index);
 ND4J_EXPORT Nd4jLong getRandomGeneratorRelativeLong(OpaqueRandomGenerator* ptr, Nd4jLong index);
 ND4J_EXPORT void deleteRandomGenerator(OpaqueRandomGenerator* ptr);

libnd4j/include/legacy/cpu/NativeOps.cpp

@@ -2832,6 +2832,14 @@ void setRandomGeneratorStates(sd::graph::RandomGenerator* ptr, Nd4jLong rootSeed
     ptr->setStates(rootSeed, nodeSeed);
 }
 
+float getRandomGeneratorRelativeFloat(sd::graph::RandomGenerator* ptr, Nd4jLong index) {
+  return ptr->relativeT<float>(index);
+}
+
+double getRandomGeneratorRelativeDouble(sd::graph::RandomGenerator* ptr, Nd4jLong index) {
+  return ptr->relativeT<double>(index);
+}
+
 int getRandomGeneratorRelativeInt(sd::graph::RandomGenerator* ptr, Nd4jLong index) {
     return ptr->relativeInt(index);
 }

libnd4j/include/legacy/cuda/NativeOps.cu

@@ -3515,6 +3515,14 @@ void setRandomGeneratorStates(sd::graph::RandomGenerator* ptr, Nd4jLong rootSeed
     ptr->setStates(rootSeed, nodeSeed);
 }
 
+float getRandomGeneratorRelativeFloat(sd::graph::RandomGenerator* ptr, Nd4jLong index) {
+  return ptr->relativeT<float>(index);
+}
+
+double getRandomGeneratorRelativeDouble(sd::graph::RandomGenerator* ptr, Nd4jLong index) {
+  return ptr->relativeT<double>(index);
+}
+
 int getRandomGeneratorRelativeInt(sd::graph::RandomGenerator* ptr, Nd4jLong index) {
     return ptr->relativeInt(index);
 }

libnd4j/include/ops/declarable/generic/images/resize_images.cpp

@@ -81,7 +81,6 @@ namespace sd {
         }
 
         DECLARE_SHAPE_FN(resize_images) {
-            auto shapeList = SHAPELIST(); 
             auto in = inputShape->at(0);
 
             Nd4jLong* outputShape;

libnd4j/include/types/u32.h

@@ -0,0 +1,40 @@
+/*******************************************************************************
+ * 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 raver119@protonmail.com
+//
+#ifndef SD_U32_H
+#define SD_U32_H
+
+#include <cstdint>
+#include <system/pointercast.h>
+
+
+namespace sd {
+    union u32 {
+        bool _bool;
+        int8_t _s8;
+        uint8_t _u8;
+        int16_t _s16;
+        uint16_t _u16;
+        int32_t _s32;
+        uint32_t _u32;
+        float _f32;
+    };
+}
+
+#endif

libnd4j/tests_cpu/layers_tests/CMakeLists.txt

@@ -45,7 +45,7 @@ if (APPLE)
     set(CMAKE_CXX_FLAGS  " -fPIC -D__APPLE_OS__=true")
 elseif(WIN32)
     if (SD_CPU)
-        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -fPIC -march=native -mtune=native -O3")
+        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -fPIC -mmmx -msse -msse2 -msse3 -mssse3 -msse4.1 -msse4.2 -msse4 -mavx -mavx2 -O3")
     endif()
 
 	if (SD_CPU AND LINUX)
@@ -61,7 +61,7 @@ else()
     endif()
 
     if (SD_CPU AND SD_SANITIZE)
-        set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS} -fsanitize=address")
+        set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -fsanitize=address")
     else()
         # CUDA?
     endif()

libnd4j/tests_cpu/layers_tests/RNGTests.cpp

@@ -49,8 +49,8 @@ public:
         //_bufferB = new Nd4jLong[100000];
         //_rngA = (sd::random::RandomBuffer *) initRandom(nullptr, _seed, 100000, (Nd4jPointer) _bufferA);
         //_rngB = (sd::random::RandomBuffer *) initRandom(nullptr, _seed, 100000, (Nd4jPointer) _bufferB);
-        _rngA.setStates(_seed, _seed);
+        _rngA.setStates(_seed * 0xDEADBEEF * 13, _seed * 0xDEADBEEF * 7);
-        _rngB.setStates(_seed, _seed);
+        _rngB.setStates(_seed * 0xDEADBEEF * 13, _seed * 0xDEADBEEF * 7);
         nexp0->assign(-1.0f);
         nexp1->assign(-2.0f);
         nexp2->assign(-3.0f);
@@ -204,6 +204,9 @@ TEST_F(RNGTests, Test_Uniform_1) {
     RandomLauncher::fillUniform(LaunchContext::defaultContext(), _rngA, &x0, 1.0f, 2.0f);
     RandomLauncher::fillUniform(LaunchContext::defaultContext(), _rngB, &x1, 1.0f, 2.0f);
 
+    x0.printLinearBuffer();
+    x1.printLinearBuffer();
+
     ASSERT_TRUE(x0.equalsTo(&x1));
 
     ASSERT_FALSE(x0.equalsTo(nexp0));
@@ -212,10 +215,82 @@ TEST_F(RNGTests, Test_Uniform_1) {
 
     for (int e = 0; e < x0.lengthOf(); e++) {
         float v = x0.e<float>(e);
+        nd4j_printf("%f\n", v);
         ASSERT_TRUE(v >= 1.0f && v <= 2.0f);
     }
 }
 
+TEST_F(RNGTests, Test_Uniform_10) {
+  auto x = NDArrayFactory::create<float>('c', {10000, 10000});
+  auto z = NDArrayFactory::create<float>(0.0f);
+
+  RandomLauncher::fillUniform(LaunchContext::defaultContext(), _rngA, &x, 0.0f, 1.0f);
+
+  sd::ops::reduce_max op;
+  auto status = op.execute({&x}, {&z});
+  ASSERT_EQ(Status::OK(), status);
+
+  ASSERT_LT(z.t<float>(0), 1.0f);
+}
+
+TEST_F(RNGTests, Test_Uniform_10_double) {
+  auto x = NDArrayFactory::create<double>('c', {10000, 10000});
+  auto z = NDArrayFactory::create<double>(0.0f);
+
+  RandomLauncher::fillUniform(LaunchContext::defaultContext(), _rngA, &x, 0.0f, 1.0f);
+
+  sd::ops::reduce_max op;
+  auto status = op.execute({&x}, {&z});
+  ASSERT_EQ(Status::OK(), status);
+
+  ASSERT_LT(z.t<double>(0), 1.0);
+}
+
+TEST_F(RNGTests, Test_Uniform_11) {
+  uint32_t max = 0;
+  for (int e = 0; e < 100000000; e++) {
+    auto v = _rngA.xoroshiro32(e) >> 8;
+    if (v > max)
+      max = v;
+  }
+
+  nd4j_printf("Max value: %i\n", (int) max);
+}
+
+TEST_F(RNGTests, Test_Uniform_12) {
+  float max = -std::numeric_limits<float>::infinity();
+  float min = std::numeric_limits<float>::infinity();
+  for (int e = 0; e < 100000000; e++) {
+    auto v = _rngA.relativeT<float>(e);
+    if (v > max)
+      max = v;
+
+    if (v < min)
+      min = v;
+  }
+
+  nd4j_printf("Max value: %.8f; Min value: %.8f\n", (float) max, (float) min);
+  ASSERT_LT(max, 1.0f);
+  ASSERT_GE(min, 0.0);
+}
+
+TEST_F(RNGTests, Test_Uniform_13) {
+  double max = -std::numeric_limits<double>::infinity();
+  double min = std::numeric_limits<double>::infinity();
+  for (int e = 0; e < 100000000; e++) {
+    auto v = _rngA.relativeT<double>(e);
+    if (v > max)
+      max = v;
+
+    if (v < min)
+      min = v;
+  }
+
+  nd4j_printf("Max value: %.8f; Min value: %.8f\n", (float) max, (float) min);
+  ASSERT_LT(max, 1.0);
+  ASSERT_GE(min, 0.0);
+}
+
 TEST_F(RNGTests, Test_Uniform_3) {
     auto x0 = NDArrayFactory::create<double>('c', {1000000});
 
@@ -258,8 +333,8 @@ TEST_F(RNGTests, Test_Gaussian_1) {
 }
 
 TEST_F(RNGTests, Test_Gaussian_21) {
-    auto x0 = NDArrayFactory::create<float>('c', {10, 10});
+    auto x0 = NDArrayFactory::create<float>('c', {1000, 1000});
-    auto x1 = NDArrayFactory::create<float>('c', {10, 10});
+    auto x1 = NDArrayFactory::create<float>('c', {1000, 1000});
 
     RandomLauncher::fillGaussian(LaunchContext::defaultContext(), _rngA, &x0, 0.0f, 1.0f);
     RandomLauncher::fillGaussian(LaunchContext::defaultContext(), _rngB, &x1, 0.0f, 1.0f);
@@ -983,6 +1058,26 @@ TEST_F(RNGTests, Test_UniformDistribution_04) {
 
 }
 
+TEST_F(RNGTests, Test_UniformDistribution_05) {
+    auto x = NDArrayFactory::create<Nd4jLong>('c', {2}, {10000, 10000});
+    auto al = NDArrayFactory::create<float>(0.f);
+    auto be = NDArrayFactory::create<float>(1.f);
+    auto exp0 = NDArrayFactory::create<float>('c', {10000, 10000});
+
+
+    sd::ops::randomuniform op;
+    auto result = op.evaluate({&x, &al, &be}, {}, {},{}, {DataType::FLOAT32});
+    ASSERT_EQ(Status::OK(), result.status());
+
+    auto z = result.at(0);
+    ASSERT_TRUE(exp0.isSameShape(z));
+    ASSERT_FALSE(exp0.equalsTo(z));
+
+    sd::ops::reduce_max checkOp;
+    auto checkResult = checkOp.evaluate({z});
+    checkResult[0]->printIndexedBuffer("Max on uniform with 0 to 1 on 100M cases is");
+}
+
 namespace sd {
     namespace tests {
         static void fillList(Nd4jLong seed, int numberOfArrays, std::vector<Nd4jLong> &shape, std::vector<NDArray*> &list, sd::graph::RandomGenerator *rng) {

libnd4j/tests_cpu/layers_tests/StringTests.cpp

@@ -25,6 +25,8 @@
 #include <array/NDArrayFactory.h>
 #include "testlayers.h"
 #include <graph/Stash.h>
+#include <helpers/BitwiseUtils.h>
+#include <bitset>
 
 using namespace sd;
 
@@ -863,3 +865,13 @@ TEST_F(StringTests, Basic_cast_UTF8toUTF32) {
     ASSERT_EQ(u8, z0);
     ASSERT_EQ(u32, z1);
 }
+
+TEST_F(StringTests, test_bit_string_1) {
+  // check bits -> vector conversion first
+  auto vec = BitwiseUtils::valueBits(1);
+
+  // check bits -> string conversion next;
+  auto str = StringUtils::bitsToString(1);
+  ASSERT_EQ(32, str.length());
+  ASSERT_EQ(std::string("00000000000000000000000000000001"), str);
+}

nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/nativeblas/NativeOps.java

@@ -1154,6 +1154,8 @@ public interface NativeOps {
     long getRandomGeneratorRootState(OpaqueRandomGenerator ptr);
     long getRandomGeneratorNodeState(OpaqueRandomGenerator ptr);
     void setRandomGeneratorStates(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long rootSeed/*=0*/, @Cast("Nd4jLong") long nodeSeed/*=0*/);
+    float getRandomGeneratorRelativeFloat(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
+    double getRandomGeneratorRelativeDouble(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
     int getRandomGeneratorRelativeInt(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
     long getRandomGeneratorRelativeLong(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
     void deleteRandomGenerator(OpaqueRandomGenerator ptr);

nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/rng/NativeRandom.java

@@ -107,14 +107,10 @@ public abstract class NativeRandom implements Random {
     }
 
     @Override
-    public float nextFloat() {
+    public abstract float nextFloat();
-        return (float) nextInt() / (float) Integer.MAX_VALUE;
-    }
 
     @Override
-    public double nextDouble() {
+    public abstract double nextDouble();
-        return (double) nextInt() / (double) Integer.MAX_VALUE;
-    }
 
     @Override
     public double nextGaussian() {

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/rng/CudaNativeRandom.java

@@ -81,6 +81,16 @@ public class CudaNativeRandom extends NativeRandom {
         return seed;
     }
 
+    @Override
+    public float nextFloat() {
+        return nativeOps.getRandomGeneratorRelativeFloat((OpaqueRandomGenerator)statePointer, currentPosition.getAndIncrement());
+    }
+
+    @Override
+    public double nextDouble() {
+        return nativeOps.getRandomGeneratorRelativeDouble((OpaqueRandomGenerator)statePointer, currentPosition.getAndIncrement());
+    }
+
     @Override
     public int nextInt() {
         return nativeOps.getRandomGeneratorRelativeInt((OpaqueRandomGenerator)statePointer, currentPosition.getAndIncrement());

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java

@@ -3098,6 +3098,8 @@ public native @Cast("Nd4jLong") long getRandomGeneratorRootState(OpaqueRandomGen
 public native @Cast("Nd4jLong") long getRandomGeneratorNodeState(OpaqueRandomGenerator ptr);
 public native void setRandomGeneratorStates(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long rootSeed/*=0*/, @Cast("Nd4jLong") long nodeSeed/*=0*/);
 public native void setRandomGeneratorStates(OpaqueRandomGenerator ptr);
+public native float getRandomGeneratorRelativeFloat(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
+public native double getRandomGeneratorRelativeDouble(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
 public native int getRandomGeneratorRelativeInt(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
 public native @Cast("Nd4jLong") long getRandomGeneratorRelativeLong(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
 public native void deleteRandomGenerator(OpaqueRandomGenerator ptr);
@@ -5048,6 +5050,7 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
 // #include <array/DataTypeUtils.h>
 // #include <helpers/logger.h>
 // #include <stdexcept>
+// #include <math/templatemath.h>
 
 // #ifdef __CUDACC__
 // #endif
@@ -5064,6 +5067,8 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
                 return (RandomGenerator)super.position(position);
             }
         
+            public native @Cast("uint32_t") int xoroshiro32(@Cast("uint64_t") long index);
+            public native @Cast("uint64_t") long xoroshiro64(@Cast("uint64_t") long index);
             public RandomGenerator(@Cast("Nd4jLong") long rootSeed/*=0*/, @Cast("Nd4jLong") long nodeSeed/*=0*/) { super((Pointer)null); allocate(rootSeed, nodeSeed); }
             private native void allocate(@Cast("Nd4jLong") long rootSeed/*=0*/, @Cast("Nd4jLong") long nodeSeed/*=0*/);
             public RandomGenerator() { super((Pointer)null); allocate(); }
@@ -5094,7 +5099,7 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
             public native int relativeInt(@Cast("Nd4jLong") long index);
             public native @Cast("Nd4jLong") long relativeLong(@Cast("Nd4jLong") long index);
 
-            public native void rewindH(@Cast("Nd4jLong") long steps);
+            public native void rewindH(@Cast("uint64_t") long steps);
 
             /**
              * These methods set up only node states, with non-changed root ones
@@ -5131,6 +5136,10 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
 
         
 
+        
+
+        
+
 
         
 
@@ -5141,6 +5150,8 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
 
         @Namespace("sd::graph") public static native @Cast("uint64_t") long rotl(@Cast("const uint64_t") long x, int k);
 
+        @Namespace("sd::graph") public static native @Cast("uint32_t") int next(@Cast("uint32_t") int s0, @Cast("uint32_t") int s1, @Cast("uint32_t") int s2, @Cast("uint32_t") int s3);
+
         
 
         

nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/linalg/cpu/nativecpu/rng/CpuNativeRandom.java

@@ -75,6 +75,16 @@ public class CpuNativeRandom extends NativeRandom {
         return nativeOps.getRandomGeneratorRelativeInt((OpaqueRandomGenerator)statePointer, currentPosition.getAndIncrement());
     }
 
+    @Override
+    public float nextFloat() {
+        return nativeOps.getRandomGeneratorRelativeFloat((OpaqueRandomGenerator)statePointer, currentPosition.getAndIncrement());
+    }
+
+    @Override
+    public double nextDouble() {
+        return nativeOps.getRandomGeneratorRelativeDouble((OpaqueRandomGenerator)statePointer, currentPosition.getAndIncrement());
+    }
+
     @Override
     public long nextLong() {
         return nativeOps.getRandomGeneratorRelativeLong((OpaqueRandomGenerator)statePointer, currentPosition.getAndIncrement());

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

@@ -3102,6 +3102,8 @@ public native @Cast("Nd4jLong") long getRandomGeneratorRootState(OpaqueRandomGen
 public native @Cast("Nd4jLong") long getRandomGeneratorNodeState(OpaqueRandomGenerator ptr);
 public native void setRandomGeneratorStates(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long rootSeed/*=0*/, @Cast("Nd4jLong") long nodeSeed/*=0*/);
 public native void setRandomGeneratorStates(OpaqueRandomGenerator ptr);
+public native float getRandomGeneratorRelativeFloat(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
+public native double getRandomGeneratorRelativeDouble(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
 public native int getRandomGeneratorRelativeInt(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
 public native @Cast("Nd4jLong") long getRandomGeneratorRelativeLong(OpaqueRandomGenerator ptr, @Cast("Nd4jLong") long index);
 public native void deleteRandomGenerator(OpaqueRandomGenerator ptr);
@@ -5052,6 +5054,7 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
 // #include <array/DataTypeUtils.h>
 // #include <helpers/logger.h>
 // #include <stdexcept>
+// #include <math/templatemath.h>
 
 // #ifdef __CUDACC__
 // #endif
@@ -5068,6 +5071,8 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
                 return (RandomGenerator)super.position(position);
             }
         
+            public native @Cast("uint32_t") int xoroshiro32(@Cast("uint64_t") long index);
+            public native @Cast("uint64_t") long xoroshiro64(@Cast("uint64_t") long index);
             public RandomGenerator(@Cast("Nd4jLong") long rootSeed/*=0*/, @Cast("Nd4jLong") long nodeSeed/*=0*/) { super((Pointer)null); allocate(rootSeed, nodeSeed); }
             private native void allocate(@Cast("Nd4jLong") long rootSeed/*=0*/, @Cast("Nd4jLong") long nodeSeed/*=0*/);
             public RandomGenerator() { super((Pointer)null); allocate(); }
@@ -5098,7 +5103,7 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
             public native int relativeInt(@Cast("Nd4jLong") long index);
             public native @Cast("Nd4jLong") long relativeLong(@Cast("Nd4jLong") long index);
 
-            public native void rewindH(@Cast("Nd4jLong") long steps);
+            public native void rewindH(@Cast("uint64_t") long steps);
 
             /**
              * These methods set up only node states, with non-changed root ones
@@ -5135,6 +5140,10 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
 
         
 
+        
+
+        
+
 
         
 
@@ -5145,6 +5154,8 @@ public native @Cast("bool") boolean isOptimalRequirementsMet();
 
         @Namespace("sd::graph") public static native @Cast("uint64_t") long rotl(@Cast("const uint64_t") long x, int k);
 
+        @Namespace("sd::graph") public static native @Cast("uint32_t") int next(@Cast("uint32_t") int s0, @Cast("uint32_t") int s1, @Cast("uint32_t") int s2, @Cast("uint32_t") int s3);
+
         
 
         

nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/rng/RandomTests.java

@@ -1514,6 +1514,28 @@ public class RandomTests extends BaseNd4jTest {
         assertEquals(res[0], res1[0]);
     }
 
+
+    @Test
+    public void testRandom() {
+        val r1 = new java.util.Random(119);
+        val r2 = Nd4j.getRandom();
+        r2.setSeed(119);
+        float jmax = 0.0f;
+        float nmax = 0.0f;
+        for (int e = 0; e < 100_000_000; e++) {
+            val f = r1.nextFloat();
+            val n = r2.nextFloat();
+            if (f > jmax)
+                jmax = f;
+
+            if (n > nmax)
+                nmax = n;
+        }
+
+        assertTrue(jmax < 1.0);
+        assertTrue(nmax < 1.0);
+    }
+
     @Override
     public char ordering() {
         return 'c';