diff --git a/libnd4j/CMakeLists.txt b/libnd4j/CMakeLists.txt
index 949dbd542..a16c9eaf7 100755
--- a/libnd4j/CMakeLists.txt
+++ b/libnd4j/CMakeLists.txt
@@ -22,8 +22,8 @@ if (APPLE)
 elseif(WIN32)
     set(X86_BUILD true)
     if (NOT CUDA_BLAS)
-        set(CMAKE_CXX_FLAGS_RELEASE  "-O3 -fPIC -std=c++11 -fmax-errors=2 -D_RELEASE=true")
-        set(CMAKE_CXX_FLAGS_DEBUG  " -g -fPIC -std=c++11 -fmax-errors=2")
+        set(CMAKE_CXX_FLAGS_RELEASE  "-O3 -fPIC -std=c++11 -fmax-errors=2 -DINLINE_LOOPS -D_RELEASE=true")
+        set(CMAKE_CXX_FLAGS_DEBUG  " -g -fPIC -std=c++11 -DINLINE_LOOPS -fmax-errors=2")
     else()
         set(CMAKE_CXX_FLAGS_RELEASE  "-D_RELEASE=true /wd4804")
         set(CMAKE_CXX_FLAGS_DEBUG  "  /FS /EHsc /wd4661 /wd4804 /wd4267 /wd4244 /wd4251 /wd4305")
diff --git a/libnd4j/include/helpers/cpu/loops/Reduction3Loops_0.cpp b/libnd4j/include/helpers/cpu/loops/Reduction3Loops_0.cpp
index 895afccfd..16bf3b08b 100644
--- a/libnd4j/include/helpers/cpu/loops/Reduction3Loops_0.cpp
+++ b/libnd4j/include/helpers/cpu/loops/Reduction3Loops_0.cpp
@@ -29,23 +29,31 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void Reduction3Loops<X,Z>::innerloopReduce3(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, int* dims, int dimsLen, Z* extraParams) {
-         Reduction3Loops<X,Z>::template loopReduce3<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams);
+#ifndef INLINE_LOOPS
+        Reduction3Loops<X,Z>::template loopReduce3<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams);
+#endif
     }
 
     template<typename X, typename Z>
     template <typename OpType>
     void Reduction3Loops<X,Z>::innerloopReduce3All(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Z* extraParams) {
-         Reduction3Loops<X,Z>::template loopReduce3All<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams);
+#ifndef INLINE_LOOPS
+        Reduction3Loops<X,Z>::template loopReduce3All<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
     void Reduction3Loops<X, Y>::wrapper(const int opNum, X *x, Nd4jLong *xShapeInfo, X *y, Nd4jLong *yShapeInfo, Y *z, Nd4jLong *zShapeInfo, int* dims, int dimsLen, Y *extraParams) {
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce3, PARAMS(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams), REDUCE3_OPS);
+#endif
     }
 
     template<typename X, typename Y>
     void Reduction3Loops<X, Y>::wrapperAll(const int opNum, X *x, Nd4jLong *xShapeInfo, X *y, Nd4jLong *yShapeInfo, Y *z, Nd4jLong *zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Y* extraParams) {
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce3All, PARAMS(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo,  xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams), REDUCE3_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT Reduction3Loops, , LIBND4J_TYPES, FLOAT_TYPES_0);
diff --git a/libnd4j/include/helpers/cpu/loops/Reduction3Loops_1.cpp b/libnd4j/include/helpers/cpu/loops/Reduction3Loops_1.cpp
index d8c24e096..4e350ce15 100644
--- a/libnd4j/include/helpers/cpu/loops/Reduction3Loops_1.cpp
+++ b/libnd4j/include/helpers/cpu/loops/Reduction3Loops_1.cpp
@@ -29,23 +29,31 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void Reduction3Loops<X,Z>::innerloopReduce3(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, int* dims, int dimsLen, Z* extraParams) {
-         Reduction3Loops<X,Z>::template loopReduce3<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams);
+#ifndef INLINE_LOOPS
+        Reduction3Loops<X,Z>::template loopReduce3<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams);
+#endif
     }
 
     template<typename X, typename Z>
     template <typename OpType>
     void Reduction3Loops<X,Z>::innerloopReduce3All(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Z* extraParams) {
-         Reduction3Loops<X,Z>::template loopReduce3All<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams);
+#ifndef INLINE_LOOPS
+        Reduction3Loops<X,Z>::template loopReduce3All<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
     void Reduction3Loops<X, Y>::wrapper(const int opNum, X *x, Nd4jLong *xShapeInfo, X *y, Nd4jLong *yShapeInfo, Y *z, Nd4jLong *zShapeInfo, int* dims, int dimsLen, Y *extraParams) {
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce3, PARAMS(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams), REDUCE3_OPS);
+#endif
     }
 
     template<typename X, typename Y>
     void Reduction3Loops<X, Y>::wrapperAll(const int opNum, X *x, Nd4jLong *xShapeInfo, X *y, Nd4jLong *yShapeInfo, Y *z, Nd4jLong *zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Y* extraParams) {
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce3All, PARAMS(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo,  xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams), REDUCE3_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT Reduction3Loops, , LIBND4J_TYPES, FLOAT_TYPES_1);
diff --git a/libnd4j/include/helpers/cpu/loops/Reduction3Loops_2.cpp b/libnd4j/include/helpers/cpu/loops/Reduction3Loops_2.cpp
index 4ecc0e370..e869793a8 100644
--- a/libnd4j/include/helpers/cpu/loops/Reduction3Loops_2.cpp
+++ b/libnd4j/include/helpers/cpu/loops/Reduction3Loops_2.cpp
@@ -29,23 +29,31 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void Reduction3Loops<X,Z>::innerloopReduce3(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, int* dims, int dimsLen, Z* extraParams) {
-         Reduction3Loops<X,Z>::template loopReduce3<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams);
+#ifndef INLINE_LOOPS
+        Reduction3Loops<X,Z>::template loopReduce3<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams);
+#endif
     }
 
     template<typename X, typename Z>
     template <typename OpType>
     void Reduction3Loops<X,Z>::innerloopReduce3All(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Z* extraParams) {
-         Reduction3Loops<X,Z>::template loopReduce3All<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams);
+#ifndef INLINE_LOOPS
+        Reduction3Loops<X,Z>::template loopReduce3All<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
     void Reduction3Loops<X, Y>::wrapper(const int opNum, X *x, Nd4jLong *xShapeInfo, X *y, Nd4jLong *yShapeInfo, Y *z, Nd4jLong *zShapeInfo, int* dims, int dimsLen, Y *extraParams) {
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce3, PARAMS(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams), REDUCE3_OPS);
+#endif
     }
 
     template<typename X, typename Y>
     void Reduction3Loops<X, Y>::wrapperAll(const int opNum, X *x, Nd4jLong *xShapeInfo, X *y, Nd4jLong *yShapeInfo, Y *z, Nd4jLong *zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Y* extraParams) {
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce3All, PARAMS(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo,  xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams), REDUCE3_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT Reduction3Loops, , LIBND4J_TYPES, FLOAT_TYPES_2);
diff --git a/libnd4j/include/helpers/cpu/loops/Reduction3Loops_3.cpp b/libnd4j/include/helpers/cpu/loops/Reduction3Loops_3.cpp
index 218c335ca..474443fd3 100644
--- a/libnd4j/include/helpers/cpu/loops/Reduction3Loops_3.cpp
+++ b/libnd4j/include/helpers/cpu/loops/Reduction3Loops_3.cpp
@@ -29,23 +29,31 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void Reduction3Loops<X,Z>::innerloopReduce3(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, int* dims, int dimsLen, Z* extraParams) {
-         Reduction3Loops<X,Z>::template loopReduce3<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams);
+#ifndef INLINE_LOOPS
+        Reduction3Loops<X,Z>::template loopReduce3<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams);
+#endif
     }
 
     template<typename X, typename Z>
     template <typename OpType>
     void Reduction3Loops<X,Z>::innerloopReduce3All(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Z* extraParams) {
-         Reduction3Loops<X,Z>::template loopReduce3All<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams);
+#ifndef INLINE_LOOPS
+        Reduction3Loops<X,Z>::template loopReduce3All<OpType>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
     void Reduction3Loops<X, Y>::wrapper(const int opNum, X *x, Nd4jLong *xShapeInfo, X *y, Nd4jLong *yShapeInfo, Y *z, Nd4jLong *zShapeInfo, int* dims, int dimsLen, Y *extraParams) {
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce3, PARAMS(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dims, dimsLen, extraParams), REDUCE3_OPS);
+#endif
     }
 
     template<typename X, typename Y>
     void Reduction3Loops<X, Y>::wrapperAll(const int opNum, X *x, Nd4jLong *xShapeInfo, X *y, Nd4jLong *yShapeInfo, Y *z, Nd4jLong *zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Y* extraParams) {
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce3All, PARAMS(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo,  xTadShapeInfo, xTadOffsets, yTadShapeInfo, yTadOffsets, extraParams), REDUCE3_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT Reduction3Loops, , LIBND4J_TYPES, FLOAT_TYPES_3);
diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops.hpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops.hpp
index 4a223a0f2..0709e5f3c 100644
--- a/libnd4j/include/helpers/cpu/loops/ReductionLoops.hpp
+++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops.hpp
@@ -19,3 +19,4 @@
 //
 
 #include <helpers/Loops.h>
+#include <op_boilerplate.h>
diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_bool.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_bool.cpp
index 35ae99afb..3d7a85eff 100644
--- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_bool.cpp
+++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_bool.cpp
@@ -27,7 +27,9 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void ReductionBoolLoops<X, Z>::innerloopReduce(X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams) {
+#ifndef INLINE_LOOPS
         ReductionLoops<X,Z,X>::template loopReduce<OpType>(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
@@ -35,7 +37,9 @@ namespace nd4j {
                                             Nd4jLong *zShapeInfo, Nd4jLong *tadShapeInfo,
                                             Nd4jLong *tadOffsets,
                                             X *extraParams) {
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams), REDUCE_BOOL_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReductionBoolLoops, , LIBND4J_TYPES, BOOL_TYPES);
diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_0.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_0.cpp
index c7b1f6ff8..f545c8c83 100644
--- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_0.cpp
+++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_0.cpp
@@ -29,15 +29,18 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void ReductionFloatLoops<X, Z>::innerloopReduce(X * x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, Z* extraParams) {
+#ifndef INLINE_LOOPS
         ReductionLoops<X,Z,Z>::template loopReduce<OpType>(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
     void ReductionFloatLoops<X, Y>::wrapper(const int opNum, X *x, Nd4jLong *xShapeInfo, Y *z,
                                                   Nd4jLong *zShapeInfo, Nd4jLong *tadShapeInfo,
                                                   Nd4jLong *tadOffsets, Y *extraParams) {
-
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams ), REDUCE_FLOAT_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReductionFloatLoops, , LIBND4J_TYPES, FLOAT_TYPES_0);
diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_1.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_1.cpp
index 76c1141bf..fa52015ca 100644
--- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_1.cpp
+++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_1.cpp
@@ -29,15 +29,18 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void ReductionFloatLoops<X, Z>::innerloopReduce(X * x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, Z* extraParams) {
+#ifndef INLINE_LOOPS
         ReductionLoops<X,Z,Z>::template loopReduce<OpType>(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
     void ReductionFloatLoops<X, Y>::wrapper(const int opNum, X *x, Nd4jLong *xShapeInfo, Y *z,
                                                   Nd4jLong *zShapeInfo, Nd4jLong *tadShapeInfo,
                                                   Nd4jLong *tadOffsets, Y *extraParams) {
-
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams ), REDUCE_FLOAT_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReductionFloatLoops, , LIBND4J_TYPES, FLOAT_TYPES_1);
diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_2.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_2.cpp
index 7288816ad..eb144fcc6 100644
--- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_2.cpp
+++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_2.cpp
@@ -29,15 +29,18 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void ReductionFloatLoops<X, Z>::innerloopReduce(X * x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, Z* extraParams) {
+#ifndef INLINE_LOOPS
         ReductionLoops<X,Z,Z>::template loopReduce<OpType>(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
     void ReductionFloatLoops<X, Y>::wrapper(const int opNum, X *x, Nd4jLong *xShapeInfo, Y *z,
                                                   Nd4jLong *zShapeInfo, Nd4jLong *tadShapeInfo,
                                                   Nd4jLong *tadOffsets, Y *extraParams) {
-
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams ), REDUCE_FLOAT_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReductionFloatLoops, , LIBND4J_TYPES, FLOAT_TYPES_2);
diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_3.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_3.cpp
index 251624076..d2991b51b 100644
--- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_3.cpp
+++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_float_3.cpp
@@ -29,15 +29,18 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void ReductionFloatLoops<X, Z>::innerloopReduce(X * x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, Z* extraParams) {
+#ifndef INLINE_LOOPS
         ReductionLoops<X,Z,Z>::template loopReduce<OpType>(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
     void ReductionFloatLoops<X, Y>::wrapper(const int opNum, X *x, Nd4jLong *xShapeInfo, Y *z,
                                                   Nd4jLong *zShapeInfo, Nd4jLong *tadShapeInfo,
                                                   Nd4jLong *tadOffsets, Y *extraParams) {
-
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams ), REDUCE_FLOAT_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReductionFloatLoops, , LIBND4J_TYPES, FLOAT_TYPES_3);
diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_long.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_long.cpp
index a6dd992c6..04a3d8559 100644
--- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_long.cpp
+++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_long.cpp
@@ -34,15 +34,18 @@ namespace nd4j {
     template<typename X, typename Z>
     template <typename OpType>
     void ReductionLongLoops<X, Z>::innerloopReduce(X * x, Nd4jLong* xShapeInfo, Z *z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams) {
+#ifndef INLINE_LOOPS
         ReductionLoops<X,Z,X>::template loopReduce<OpType>(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams);
+#endif
     }
 
     template<typename X, typename Y>
     void ReductionLongLoops<X, Y>::wrapper(const int opNum, X *x, Nd4jLong *xShapeInfo, Y *z,
                                             Nd4jLong *zShapeInfo, Nd4jLong *tadShapeInfo,
                                             Nd4jLong *tadOffsets, X *extraParams) {
-
+#ifndef INLINE_LOOPS
         DISPATCH_BY_OPNUM_TT(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams ), REDUCE_LONG_OPS);
+#endif
     }
 
     BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT ReductionLongLoops, , LIBND4J_TYPES, LONG_TYPES);
diff --git a/libnd4j/include/helpers/cpu/loops/ReductionLoops_same.cpp b/libnd4j/include/helpers/cpu/loops/ReductionLoops_same.cpp
index 623d97e79..9932b04c5 100644
--- a/libnd4j/include/helpers/cpu/loops/ReductionLoops_same.cpp
+++ b/libnd4j/include/helpers/cpu/loops/ReductionLoops_same.cpp
@@ -27,7 +27,9 @@ namespace nd4j {
     template<typename X>
     template <typename OpType>
     void ReductionSameLoops<X>::innerloopReduce(X* x, Nd4jLong* xShapeInfo, X* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams) {
+#ifndef INLINE_LOOPS
         ReductionLoops<X,X,X>::template loopReduce<OpType>(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams);
+#endif
     }
 
     template<typename X>
@@ -35,11 +37,13 @@ namespace nd4j {
                                            Nd4jLong *zShapeInfo, Nd4jLong *tadShapeInfo,
                                            Nd4jLong *tadOffsets,
                                            X *vextraParams) {
+#ifndef INLINE_LOOPS
         auto x = reinterpret_cast<X *>(vx);
         auto z = reinterpret_cast<X *>(vz);
         auto extraParams = reinterpret_cast<X *>(vextraParams);
 
         DISPATCH_BY_OPNUM_T(innerloopReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams), REDUCE_SAME_OPS);
+#endif
     }
 
     BUILD_SINGLE_TEMPLATE(template class ReductionSameLoops, , LIBND4J_TYPES);