diff --git a/Build.py b/Build.py
index c4ec4f6..06c1255 100644
--- a/Build.py
+++ b/Build.py
@@ -89,6 +89,7 @@ def main():
     # Provisional extensions
     parser.add_option("--tiling", dest="tiling", help="Add -DOPENVX_USE_TILING=ON to support the tiling extension", default=False, action='store_true')
     parser.add_option("--s16", dest="s16", help="Add -DOPENVX_USE_S16=ON to have an extended support for S16", default=False, action='store_true')
+    parser.add_option("--userdataobj", dest="userdataobj", help="Add -DOPENVX_USE_USER_DATA_OBJECT=ON to support user data object extension", default=False, action='store_true')
     # Experimental features
     parser.add_option("--f16", dest="f16", help="Add -DEXPERIMENTAL_PLATFORM_SUPPORTS_16_FLOAT=ON to support VX_TYPE_FLOAT16", default=False, action='store_true')
     parser.add_option("--venum", dest="venum", help="Add -DEXPERIMENTAL_USE_VENUM=ON to build also raspberrypi 3B+ Neon target[Default False]", default=False, action='store_true')
@@ -232,6 +233,8 @@ def main():
         cmd += ['-DEXPERIMENTAL_USE_VENUM=ON']
     if options.opencl:
         cmd += ['-DEXPERIMENTAL_USE_OPENCL=ON']
+    if options.userdataobj:
+        cmd += ['-DOPENVX_USE_USER_DATA_OBJECT=ON']
     cmd = ' '.join(cmd)
 
     print( "" )
diff --git a/cmake_utils/CMakeFuncs.txt b/cmake_utils/CMakeFuncs.txt
index f90f2a8..fc3f5ef 100644
--- a/cmake_utils/CMakeFuncs.txt
+++ b/cmake_utils/CMakeFuncs.txt
@@ -94,7 +94,9 @@ endfunction ( FIND_SOURCES )
 #
 function ( FIND_NUM_PROCESSORS )
   set(PROCESSOR_COUNT_T 0)
-  if ( WIN32 )
+  if(CMAKE_CROSSCOMPILING)
+    set(PROCESSOR_COUNT_T "1")
+  elseif ( WIN32 )
     set(PROCESSOR_COUNT_T "$ENV{NUMBER_OF_PROCESSORS}")
   else ( WIN32 )
 	set(CPUINFO_FILE "/proc/cpuinfo")
diff --git a/cmake_utils/CMake_linux_tools.cmake b/cmake_utils/CMake_linux_tools.cmake
index 02df94b..3cd2688 100644
--- a/cmake_utils/CMake_linux_tools.cmake
+++ b/cmake_utils/CMake_linux_tools.cmake
@@ -29,7 +29,9 @@ endif (CYGWIN)
 
 if(BUILD_X64)
   if (EXPERIMENTAL_USE_VENUM OR OPENVX_USE_TILING OR EXPERIMENTAL_USE_OPENCL)
-    set(ARCH_BIT "-mfpu=neon")
+    if (NOT (CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64"))
+      set(ARCH_BIT "-mfpu=neon")
+    endif()
   else ()
     set(ARCH_BIT -m64 )
   endif(EXPERIMENTAL_USE_VENUM OR OPENVX_USE_TILING OR EXPERIMENTAL_USE_OPENCL)
@@ -38,7 +40,9 @@ else()
     # architecture will be according to ATOM
     set(ARCH_BIT -m32 )
   elseif (EXPERIMENTAL_USE_VENUM OR OPENVX_USE_TILING OR EXPERIMENTAL_USE_OPENCL)
-    set(ARCH_BIT "-mfpu=neon")
+    if (NOT (CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64"))
+      set(ARCH_BIT "-mfpu=neon")
+    endif()
   else ()
     # need to force a more modern architecture than the degault m32 (i386).
     set(ARCH_BIT "-m32 -march=core2" )
diff --git a/cmake_utils/CMake_windows_tools.cmake b/cmake_utils/CMake_windows_tools.cmake
index a7457fb..e2b4e7c 100644
--- a/cmake_utils/CMake_windows_tools.cmake
+++ b/cmake_utils/CMake_windows_tools.cmake
@@ -18,7 +18,7 @@
 
 # Compiler switches that CANNOT be modified during makefile generation
 set (ADD_C_FLAGS         "/Oi -D WINDOWS_ENABLE_CPLUSPLUS /GS")
-set (ADD_C_FLAGS_DEBUG   "-D _DEBUG /RTC1 /MTd")  #/MTd /Gm
+set (ADD_C_FLAGS_DEBUG   "-D _DEBUG /RTC1 /MTd /bigobj")  #/MTd /Gm
 set (ADD_C_FLAGS_RELEASE "/Zi /Gy -D NDEBUG /MT")# /Ob0") #/GL") #MT
 
 # Compiler switches that CAN be modified during makefile generation and configuration-independent
diff --git a/kernels/c_model/c_khr_nn.c b/kernels/c_model/c_khr_nn.c
index 7e29227..00ad99a 100644
--- a/kernels/c_model/c_khr_nn.c
+++ b/kernels/c_model/c_khr_nn.c
@@ -508,7 +508,7 @@ void PoolingKernelImpl(
         if (!max_pooling)
         {
             //result = conversion_24_8(result / (int16_t)(size_x * size_y));
-            result = CLAMP(result / (size_x * size_y), getMinValue(fmt), getMaxValue(fmt));
+            result = CLAMP(result / (int32_t)(size_x * size_y), getMinValue(fmt), getMaxValue(fmt));
         }
 
         const size_t output_byte_offset =
diff --git a/kernels/extras/extras_k.h b/kernels/extras/extras_k.h
index 8cb02b9..9e3b662 100644
--- a/kernels/extras/extras_k.h
+++ b/kernels/extras/extras_k.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _VX_EXTRAS_K_H_
-#define _VX_EXTRAS_K_H_
+#ifndef VX_EXTRAS_K_H_
+#define VX_EXTRAS_K_H_
 
 #include <VX/vx.h>
 #include <VX/vx_helper.h>
diff --git a/sample/framework/vx_graph.c b/sample/framework/vx_graph.c
index 8651d14..73d7479 100644
--- a/sample/framework/vx_graph.c
+++ b/sample/framework/vx_graph.c
@@ -655,6 +655,26 @@ void ownDestructGraph(vx_reference ref)
         }
         ownRemoveNodeInt(&graph->nodes[0]);
     }
+    /* Release virtual objects scoped to this graph */
+    {
+        vx_context context = graph->base.context;
+        vx_uint32 r;
+        for (r = 0u; r < context->num_references; r++)
+        {
+            vx_reference virt_ref = context->reftable[r];
+            if (virt_ref != NULL && virt_ref->is_virtual == vx_true_e && virt_ref->scope == (vx_reference)graph)
+            {
+                if (virt_ref->external_count > 0)
+                {
+                    ownReleaseReferenceInt(&virt_ref, virt_ref->type, VX_EXTERNAL, NULL);
+                }
+                else if (virt_ref->internal_count > 0)
+                {
+                    ownDecrementReference(virt_ref, VX_INTERNAL);
+                }
+            }
+        }
+    }
     // execution lock?
     ownDestroySem(&graph->lock);
 }
diff --git a/sample/framework/vx_scalar.c b/sample/framework/vx_scalar.c
index f68886c..d1e20fe 100644
--- a/sample/framework/vx_scalar.c
+++ b/sample/framework/vx_scalar.c
@@ -447,63 +447,70 @@ VX_API_ENTRY vx_status VX_API_CALL vxReadScalarValue(vx_scalar scalar, void *ptr
         return VX_ERROR_INVALID_PARAMETERS;
 
     ownSemWait(&scalar->base.lock);
-    vxPrintScalarValue(scalar);
-    switch (scalar->data_type)
+    if (scalar->data_addr != NULL && scalar->data_len != 0)
     {
-        case VX_TYPE_CHAR:
-            *(vx_char *)ptr = scalar->data.chr;
-            break;
-        case VX_TYPE_INT8:
-            *(vx_int8 *)ptr = scalar->data.s08;
-            break;
-        case VX_TYPE_UINT8:
-            *(vx_uint8 *)ptr = scalar->data.u08;
-            break;
-        case VX_TYPE_INT16:
-            *(vx_int16 *)ptr = scalar->data.s16;
-            break;
-        case VX_TYPE_UINT16:
-            *(vx_uint16 *)ptr = scalar->data.u16;
-            break;
-        case VX_TYPE_INT32:
-            *(vx_int32 *)ptr = scalar->data.s32;
-            break;
-        case VX_TYPE_UINT32:
-            *(vx_uint32 *)ptr = scalar->data.u32;
-            break;
-        case VX_TYPE_INT64:
-            *(vx_int64 *)ptr = scalar->data.s64;
-            break;
-        case VX_TYPE_UINT64:
-            *(vx_uint64 *)ptr = scalar->data.u64;
-            break;
+        memcpy(ptr, scalar->data_addr, scalar->data_len);
+    }
+    else
+    {
+        vxPrintScalarValue(scalar);
+        switch (scalar->data_type)
+        {
+            case VX_TYPE_CHAR:
+                *(vx_char *)ptr = scalar->data.chr;
+                break;
+            case VX_TYPE_INT8:
+                *(vx_int8 *)ptr = scalar->data.s08;
+                break;
+            case VX_TYPE_UINT8:
+                *(vx_uint8 *)ptr = scalar->data.u08;
+                break;
+            case VX_TYPE_INT16:
+                *(vx_int16 *)ptr = scalar->data.s16;
+                break;
+            case VX_TYPE_UINT16:
+                *(vx_uint16 *)ptr = scalar->data.u16;
+                break;
+            case VX_TYPE_INT32:
+                *(vx_int32 *)ptr = scalar->data.s32;
+                break;
+            case VX_TYPE_UINT32:
+                *(vx_uint32 *)ptr = scalar->data.u32;
+                break;
+            case VX_TYPE_INT64:
+                *(vx_int64 *)ptr = scalar->data.s64;
+                break;
+            case VX_TYPE_UINT64:
+                *(vx_uint64 *)ptr = scalar->data.u64;
+                break;
 #if OVX_SUPPORT_HALF_FLOAT
-        case VX_TYPE_FLOAT16:
-            *(vx_float16 *)ptr = scalar->data.f16;
-            break;
+            case VX_TYPE_FLOAT16:
+                *(vx_float16 *)ptr = scalar->data.f16;
+                break;
 #endif
-        case VX_TYPE_FLOAT32:
-            *(vx_float32 *)ptr = scalar->data.f32;
-            break;
-        case VX_TYPE_FLOAT64:
-            *(vx_float64 *)ptr = scalar->data.f64;
-            break;
-        case VX_TYPE_DF_IMAGE:
-            *(vx_df_image *)ptr = scalar->data.fcc;
-            break;
-        case VX_TYPE_ENUM:
-            *(vx_enum *)ptr = scalar->data.enm;
-            break;
-        case VX_TYPE_SIZE:
-            *(vx_size *)ptr = scalar->data.size;
-            break;
-        case VX_TYPE_BOOL:
-            *(vx_bool *)ptr = scalar->data.boolean;
-            break;
-        default:
-            VX_PRINT(VX_ZONE_ERROR, "some case is not covered in %s\n", __FUNCTION__);
-            status = VX_ERROR_NOT_SUPPORTED;
-            break;
+            case VX_TYPE_FLOAT32:
+                *(vx_float32 *)ptr = scalar->data.f32;
+                break;
+            case VX_TYPE_FLOAT64:
+                *(vx_float64 *)ptr = scalar->data.f64;
+                break;
+            case VX_TYPE_DF_IMAGE:
+                *(vx_df_image *)ptr = scalar->data.fcc;
+                break;
+            case VX_TYPE_ENUM:
+                *(vx_enum *)ptr = scalar->data.enm;
+                break;
+            case VX_TYPE_SIZE:
+                *(vx_size *)ptr = scalar->data.size;
+                break;
+            case VX_TYPE_BOOL:
+                *(vx_bool *)ptr = scalar->data.boolean;
+                break;
+            default:
+                VX_PRINT(VX_ZONE_ERROR, "some case is not covered in %s\n", __FUNCTION__);
+                status = VX_ERROR_NOT_SUPPORTED;
+                break;
+        }
     }
     ownSemPost(&scalar->base.lock);
     ownReadFromReference(&scalar->base);
@@ -521,64 +528,71 @@ VX_API_ENTRY vx_status VX_API_CALL vxWriteScalarValue(vx_scalar scalar, const vo
         return VX_ERROR_INVALID_PARAMETERS;
 
     ownSemWait(&scalar->base.lock);
-    switch (scalar->data_type)
+    if (scalar->data_addr != NULL && scalar->data_len != 0)
     {
-        case VX_TYPE_CHAR:
-            scalar->data.chr = *(vx_char *)ptr;
-            break;
-        case VX_TYPE_INT8:
-            scalar->data.s08 = *(vx_int8 *)ptr;
-            break;
-        case VX_TYPE_UINT8:
-            scalar->data.u08 = *(vx_uint8 *)ptr;
-            break;
-        case VX_TYPE_INT16:
-            scalar->data.s16 = *(vx_int16 *)ptr;
-            break;
-        case VX_TYPE_UINT16:
-            scalar->data.u16 = *(vx_uint16 *)ptr;
-            break;
-        case VX_TYPE_INT32:
-            scalar->data.s32 = *(vx_int32 *)ptr;
-            break;
-        case VX_TYPE_UINT32:
-            scalar->data.u32 = *(vx_uint32 *)ptr;
-            break;
-        case VX_TYPE_INT64:
-            scalar->data.s64 = *(vx_int64 *)ptr;
-            break;
-        case VX_TYPE_UINT64:
-            scalar->data.u64 = *(vx_uint64 *)ptr;
-            break;
+        memcpy(scalar->data_addr, ptr, scalar->data_len);
+    }
+    else
+    {
+        switch (scalar->data_type)
+        {
+            case VX_TYPE_CHAR:
+                scalar->data.chr = *(vx_char *)ptr;
+                break;
+            case VX_TYPE_INT8:
+                scalar->data.s08 = *(vx_int8 *)ptr;
+                break;
+            case VX_TYPE_UINT8:
+                scalar->data.u08 = *(vx_uint8 *)ptr;
+                break;
+            case VX_TYPE_INT16:
+                scalar->data.s16 = *(vx_int16 *)ptr;
+                break;
+            case VX_TYPE_UINT16:
+                scalar->data.u16 = *(vx_uint16 *)ptr;
+                break;
+            case VX_TYPE_INT32:
+                scalar->data.s32 = *(vx_int32 *)ptr;
+                break;
+            case VX_TYPE_UINT32:
+                scalar->data.u32 = *(vx_uint32 *)ptr;
+                break;
+            case VX_TYPE_INT64:
+                scalar->data.s64 = *(vx_int64 *)ptr;
+                break;
+            case VX_TYPE_UINT64:
+                scalar->data.u64 = *(vx_uint64 *)ptr;
+                break;
 #if OVX_SUPPORT_HALF_FLOAT
-        case VX_TYPE_FLOAT16:
-            scalar->data.f16 = *(vx_float16 *)ptr;
-            break;
+            case VX_TYPE_FLOAT16:
+                scalar->data.f16 = *(vx_float16 *)ptr;
+                break;
 #endif
-        case VX_TYPE_FLOAT32:
-            scalar->data.f32 = *(vx_float32 *)ptr;
-            break;
-        case VX_TYPE_FLOAT64:
-            scalar->data.f64 = *(vx_float64 *)ptr;
-            break;
-        case VX_TYPE_DF_IMAGE:
-            scalar->data.fcc = *(vx_df_image *)ptr;
-            break;
-        case VX_TYPE_ENUM:
-            scalar->data.enm = *(vx_enum *)ptr;
-            break;
-        case VX_TYPE_SIZE:
-            scalar->data.size = *(vx_size *)ptr;
-            break;
-        case VX_TYPE_BOOL:
-            scalar->data.boolean = *(vx_bool *)ptr;
-            break;
-        default:
-            VX_PRINT(VX_ZONE_ERROR, "some case is not covered in %s\n", __FUNCTION__);
-            status = VX_ERROR_NOT_SUPPORTED;
-            break;
+            case VX_TYPE_FLOAT32:
+                scalar->data.f32 = *(vx_float32 *)ptr;
+                break;
+            case VX_TYPE_FLOAT64:
+                scalar->data.f64 = *(vx_float64 *)ptr;
+                break;
+            case VX_TYPE_DF_IMAGE:
+                scalar->data.fcc = *(vx_df_image *)ptr;
+                break;
+            case VX_TYPE_ENUM:
+                scalar->data.enm = *(vx_enum *)ptr;
+                break;
+            case VX_TYPE_SIZE:
+                scalar->data.size = *(vx_size *)ptr;
+                break;
+            case VX_TYPE_BOOL:
+                scalar->data.boolean = *(vx_bool *)ptr;
+                break;
+            default:
+                VX_PRINT(VX_ZONE_ERROR, "some case is not covered in %s\n", __FUNCTION__);
+                status = VX_ERROR_NOT_SUPPORTED;
+                break;
+        }
+        vxPrintScalarValue(scalar);
     }
-    vxPrintScalarValue(scalar);
     ownSemPost(&scalar->base.lock);
     ownWroteToReference(&scalar->base);
     return status;
diff --git a/sample/framework/vx_tensor.c b/sample/framework/vx_tensor.c
index 1e2bada..ac3357c 100644
--- a/sample/framework/vx_tensor.c
+++ b/sample/framework/vx_tensor.c
@@ -635,7 +635,7 @@ VX_API_ENTRY vx_status VX_API_CALL vxMapTensorPatch(vx_tensor tensor, vx_size nu
     stride[0] = ownSizeOfType(tensor->data_type);
     for (vx_uint32 i = 1; i < number_of_dims; i++)
     {
-        stride[i] = stride[i - 1] * (view_end[i] - view_start[i]);
+        stride[i] = stride[i - 1] * tensor->dimensions[i - 1];
     }
     //vx_map_id * map_id, vx_size * stride, void ** ptr
     size = ComputePatchSize(view_start, view_end, number_of_dims);
diff --git a/sample/framework/vx_threshold.c b/sample/framework/vx_threshold.c
index 86fe88d..22a7f16 100644
--- a/sample/framework/vx_threshold.c
+++ b/sample/framework/vx_threshold.c
@@ -422,6 +422,7 @@ VX_API_ENTRY vx_threshold VX_API_CALL vxCreateThresholdForImage(vx_context conte
         VX_PRINT(VX_ZONE_ERROR, "Invalid threshold type\n");
         vxAddLogEntry(&context->base, VX_ERROR_INVALID_TYPE, "Invalid threshold type\n");
         threshold = (vx_threshold )ownGetErrorObject(context, VX_ERROR_INVALID_TYPE);
+        return threshold;
     }
 
     if ( ((vxIsValidThresholdFormat  (input_format) == vx_false_e) &&
@@ -432,6 +433,7 @@ VX_API_ENTRY vx_threshold VX_API_CALL vxCreateThresholdForImage(vx_context conte
         VX_PRINT(VX_ZONE_ERROR, "Invalid input or output format\n");
         vxAddLogEntry(&context->base, VX_ERROR_INVALID_TYPE, "Invalid input or output format\n");
         threshold = (vx_threshold )ownGetErrorObject(context, VX_ERROR_INVALID_TYPE);
+        return threshold;
     }
 
 
diff --git a/sample/framework/vx_xml_export.c b/sample/framework/vx_xml_export.c
index 4a458c1..ae6ee5c 100644
--- a/sample/framework/vx_xml_export.c
+++ b/sample/framework/vx_xml_export.c
@@ -902,7 +902,7 @@ static vx_status vxExportToXMLThreshold(FILE* fp, vx_reference refs[], vx_uint32
     indent[i] = '\0';
 
     fprintf(fp, "%s<threshold reference=\"%u\" elemType=\"%s\" true_value=\"%d\" false_value=\"%d\"%s",
-                 indent, r, type_pairs[j].name, thresh->true_value, thresh->false_value, refNameStr);
+                 indent, r, type_pairs[j].name, thresh->true_value.S32, thresh->false_value.S32, refNameStr);
 
     if (refs[r]->is_virtual == vx_true_e) /* is not virtual in 1.0, but check anyway */
     {
diff --git a/sample/include/vx_array.h b/sample/include/vx_array.h
index 60e9db7..774e3f7 100644
--- a/sample/include/vx_array.h
+++ b/sample/include/vx_array.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_ARRAY_H_
-#define _OPENVX_INT_ARRAY_H_
+#ifndef OPENVX_INT_ARRAY_H_
+#define OPENVX_INT_ARRAY_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_context.h b/sample/include/vx_context.h
index 04e2b39..588d3e4 100644
--- a/sample/include/vx_context.h
+++ b/sample/include/vx_context.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_CONTEXT_H_
-#define _OPENVX_INT_CONTEXT_H_
+#ifndef OPENVX_INT_CONTEXT_H_
+#define OPENVX_INT_CONTEXT_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_convolution.h b/sample/include/vx_convolution.h
index 11a5bd3..44ccd48 100644
--- a/sample/include/vx_convolution.h
+++ b/sample/include/vx_convolution.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_CONVOLUTION_H_
-#define _OPENVX_INT_CONVOLUTION_H_
+#ifndef OPENVX_INT_CONVOLUTION_H_
+#define OPENVX_INT_CONVOLUTION_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_delay.h b/sample/include/vx_delay.h
index 80957c6..cdd0491 100644
--- a/sample/include/vx_delay.h
+++ b/sample/include/vx_delay.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_DELAY_H_
-#define _OPENVX_INT_DELAY_H_
+#ifndef OPENVX_INT_DELAY_H_
+#define OPENVX_INT_DELAY_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_distribution.h b/sample/include/vx_distribution.h
index f6fced3..11fec11 100644
--- a/sample/include/vx_distribution.h
+++ b/sample/include/vx_distribution.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_DISTRIBUTION_H_
-#define _OPENVX_INT_DISTRIBUTION_H_
+#ifndef OPENVX_INT_DISTRIBUTION_H_
+#define OPENVX_INT_DISTRIBUTION_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_error.h b/sample/include/vx_error.h
index 6bf1881..5850a73 100644
--- a/sample/include/vx_error.h
+++ b/sample/include/vx_error.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_ERROR_H_
-#define _OPENVX_INT_ERROR_H_
+#ifndef OPENVX_INT_ERROR_H_
+#define OPENVX_INT_ERROR_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_graph.h b/sample/include/vx_graph.h
index 5baaab6..ee91f16 100644
--- a/sample/include/vx_graph.h
+++ b/sample/include/vx_graph.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_GRAPH_H_
-#define _OPENVX_INT_GRAPH_H_
+#ifndef OPENVX_INT_GRAPH_H_
+#define OPENVX_INT_GRAPH_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_image.h b/sample/include/vx_image.h
index aef2dcb..e85a1f1 100644
--- a/sample/include/vx_image.h
+++ b/sample/include/vx_image.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_IMAGE_H_
-#define _OPENVX_INT_IMAGE_H_
+#ifndef OPENVX_INT_IMAGE_H_
+#define OPENVX_INT_IMAGE_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_import.h b/sample/include/vx_import.h
index a475c75..475c1a3 100644
--- a/sample/include/vx_import.h
+++ b/sample/include/vx_import.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_IMPORT_H_
-#define _OPENVX_INT_IMPORT_H_
+#ifndef OPENVX_INT_IMPORT_H_
+#define OPENVX_INT_IMPORT_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_inlines.c b/sample/include/vx_inlines.c
index b455f62..b8e162c 100644
--- a/sample/include/vx_inlines.c
+++ b/sample/include/vx_inlines.c
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _VX_INLINES_C_
-#define _VX_INLINES_C_
+#ifndef VX_INLINES_C_
+#define VX_INLINES_C_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_internal.h b/sample/include/vx_internal.h
index cef2dfa..0d34ca4 100644
--- a/sample/include/vx_internal.h
+++ b/sample/include/vx_internal.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _VX_INTERNAL_H_
-#define _VX_INTERNAL_H_
+#ifndef VX_INTERNAL_H_
+#define VX_INTERNAL_H_
 
 /*!
  * \file vx_internal.h
diff --git a/sample/include/vx_kernel.h b/sample/include/vx_kernel.h
index 7e61dfe..f2d3abd 100644
--- a/sample/include/vx_kernel.h
+++ b/sample/include/vx_kernel.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_KERNEL_H_
-#define _OPENVX_INT_KERNEL_H_
+#ifndef OPENVX_INT_KERNEL_H_
+#define OPENVX_INT_KERNEL_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_log.h b/sample/include/vx_log.h
index 8c532e9..870f892 100644
--- a/sample/include/vx_log.h
+++ b/sample/include/vx_log.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_LOG_H_
-#define _OPENVX_INT_LOG_H_
+#ifndef OPENVX_INT_LOG_H_
+#define OPENVX_INT_LOG_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_lut.h b/sample/include/vx_lut.h
index 35f3505..9c682ec 100644
--- a/sample/include/vx_lut.h
+++ b/sample/include/vx_lut.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_LUT_H_
-#define _OPENVX_INT_LUT_H_
+#ifndef OPENVX_INT_LUT_H_
+#define OPENVX_INT_LUT_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_matrix.h b/sample/include/vx_matrix.h
index 6d278b2..ae68f92 100644
--- a/sample/include/vx_matrix.h
+++ b/sample/include/vx_matrix.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_MATRIX_H_
-#define _OPENVX_INT_MATRIX_H_
+#ifndef OPENVX_INT_MATRIX_H_
+#define OPENVX_INT_MATRIX_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_memory.h b/sample/include/vx_memory.h
index a2be0db..e53f49c 100644
--- a/sample/include/vx_memory.h
+++ b/sample/include/vx_memory.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_MEMORY_H_
-#define _OPENVX_INT_MEMORY_H_
+#ifndef OPENVX_INT_MEMORY_H_
+#define OPENVX_INT_MEMORY_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_meta_format.h b/sample/include/vx_meta_format.h
index 443265c..38095ad 100644
--- a/sample/include/vx_meta_format.h
+++ b/sample/include/vx_meta_format.h
@@ -16,8 +16,8 @@
  */
 
 
-#ifndef _OPENVX_INT_META_FORMAT_H_
-#define _OPENVX_INT_META_FORMAT_H_
+#ifndef OPENVX_INT_META_FORMAT_H_
+#define OPENVX_INT_META_FORMAT_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_node.h b/sample/include/vx_node.h
index 72d6e73..de1b1b4 100644
--- a/sample/include/vx_node.h
+++ b/sample/include/vx_node.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_NODE_H_
-#define _OPENVX_INT_NODE_H_
+#ifndef OPENVX_INT_NODE_H_
+#define OPENVX_INT_NODE_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_object_array.h b/sample/include/vx_object_array.h
index f61d1f8..4f687c2 100644
--- a/sample/include/vx_object_array.h
+++ b/sample/include/vx_object_array.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_OBJECT_ARRAY_H_
-#define _OPENVX_INT_OBJECT_ARRAY_H_
+#ifndef OPENVX_INT_OBJECT_ARRAY_H_
+#define OPENVX_INT_OBJECT_ARRAY_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_osal.h b/sample/include/vx_osal.h
index c19c867..ce822bd 100644
--- a/sample/include/vx_osal.h
+++ b/sample/include/vx_osal.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_OSAL_H_
-#define _OPENVX_INT_OSAL_H_
+#ifndef OPENVX_INT_OSAL_H_
+#define OPENVX_INT_OSAL_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_parameter.h b/sample/include/vx_parameter.h
index c9aabbc..08e33af 100644
--- a/sample/include/vx_parameter.h
+++ b/sample/include/vx_parameter.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_PARAMETER_H_
-#define _OPENVX_INT_PARAMETER_H_
+#ifndef OPENVX_INT_PARAMETER_H_
+#define OPENVX_INT_PARAMETER_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_pyramid.h b/sample/include/vx_pyramid.h
index 8db49c7..6b50c53 100644
--- a/sample/include/vx_pyramid.h
+++ b/sample/include/vx_pyramid.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_PYRAMID_H_
-#define _OPENVX_INT_PYRAMID_H_
+#ifndef OPENVX_INT_PYRAMID_H_
+#define OPENVX_INT_PYRAMID_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_reference.h b/sample/include/vx_reference.h
index bed04db..6444db2 100644
--- a/sample/include/vx_reference.h
+++ b/sample/include/vx_reference.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_REF_H_
-#define _OPENVX_INT_REF_H_
+#ifndef OPENVX_INT_REF_H_
+#define OPENVX_INT_REF_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_remap.h b/sample/include/vx_remap.h
index 750eb14..2792e40 100644
--- a/sample/include/vx_remap.h
+++ b/sample/include/vx_remap.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_REMAP_H_
-#define _OPENVX_INT_REMAP_H_
+#ifndef OPENVX_INT_REMAP_H_
+#define OPENVX_INT_REMAP_H_
 
 #include <VX/vx.h>
 
diff --git a/sample/include/vx_sample.h b/sample/include/vx_sample.h
index f872c5c..338e878 100644
--- a/sample/include/vx_sample.h
+++ b/sample/include/vx_sample.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _VX_SAMPLE_H_
-#define _VX_SAMPLE_H_
+#ifndef VX_SAMPLE_H_
+#define VX_SAMPLE_H_
 
 #include <VX/vx.h>
 
@@ -49,5 +49,5 @@ enum _vx_sample_kernels_e {
  */
 vx_node ownCreateNodeFromGraph(vx_graph parent, vx_graph child);
 
-#endif /* _VX_SAMPLE_H_ */
+#endif /* VX_SAMPLE_H_ */
 
diff --git a/sample/include/vx_scalar.h b/sample/include/vx_scalar.h
index 12408e3..90f8a3b 100644
--- a/sample/include/vx_scalar.h
+++ b/sample/include/vx_scalar.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_SCALAR_H_
-#define _OPENVX_INT_SCALAR_H_
+#ifndef OPENVX_INT_SCALAR_H_
+#define OPENVX_INT_SCALAR_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_target.h b/sample/include/vx_target.h
index d237832..cc5001f 100644
--- a/sample/include/vx_target.h
+++ b/sample/include/vx_target.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_TARGET_H_
-#define _OPENVX_INT_TARGET_H_
+#ifndef OPENVX_INT_TARGET_H_
+#define OPENVX_INT_TARGET_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_tensor.h b/sample/include/vx_tensor.h
index c5002b2..60c556c 100644
--- a/sample/include/vx_tensor.h
+++ b/sample/include/vx_tensor.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_TENSOR_H_
-#define _OPENVX_INT_TENSOR_H_
+#ifndef OPENVX_INT_TENSOR_H_
+#define OPENVX_INT_TENSOR_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_threshold.h b/sample/include/vx_threshold.h
index 6e8219e..d83b126 100644
--- a/sample/include/vx_threshold.h
+++ b/sample/include/vx_threshold.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_threshold_H_
-#define _OPENVX_INT_threshold_H_
+#ifndef OPENVX_INT_threshold_H_
+#define OPENVX_INT_threshold_H_
 
 #include <VX/vx.h>
 #include "vx_internal.h"
diff --git a/sample/include/vx_type_pairs.h b/sample/include/vx_type_pairs.h
index 8a5b063..55466ff 100644
--- a/sample/include/vx_type_pairs.h
+++ b/sample/include/vx_type_pairs.h
@@ -15,8 +15,8 @@
  * limitations under the License.
  */
 
-#ifndef _OPENVX_INT_type_pairs_H_
-#define _OPENVX_INT_type_pairs_H_
+#ifndef OPENVX_INT_type_pairs_H_
+#define OPENVX_INT_type_pairs_H_
 
 #include <stdint.h>
 #include <VX/vx.h>
diff --git a/sample/targets/c_model/vx_pyramid.c b/sample/targets/c_model/vx_pyramid.c
index 30f0816..9f036d0 100644
--- a/sample/targets/c_model/vx_pyramid.c
+++ b/sample/targets/c_model/vx_pyramid.c
@@ -691,15 +691,95 @@ static vx_param_description_t laplacian_pyramid_kernel_params[] =
 
 static vx_status VX_CALLBACK vxLaplacianPyramidKernel(vx_node node, const vx_reference parameters[], vx_uint32 num)
 {
-    vx_status status = VX_FAILURE;
-    (void)node;
-    (void)parameters;
+    vx_status status = VX_SUCCESS;
 
     if (num == dimof(laplacian_pyramid_kernel_params))
     {
-        status = VX_SUCCESS;
+        vx_context context = vxGetContext((vx_reference)node);
+
+        vx_size lev;
+        vx_size levels = 1;
+        vx_uint32 width = 0;
+        vx_uint32 height = 0;
+        vx_uint32 level_width = 0;
+        vx_uint32 level_height = 0;
+        vx_df_image format;
+        vx_enum policy = VX_CONVERT_POLICY_SATURATE;
+        vx_border_t border;
+        vx_convolution conv = 0;
+        vx_image pyr_gauss_curr_level_filtered = 0;
+        vx_image pyr_laplacian_curr_level = 0;
+        vx_image   input = (vx_image)parameters[0];
+        vx_pyramid laplacian = (vx_pyramid)parameters[1];
+        vx_image   output = (vx_image)parameters[2];
+        vx_pyramid gaussian = 0;
+        vx_image gauss_cur = 0;
+        vx_image gauss_next = 0;
+
+        status |= vxQueryImage(input, VX_IMAGE_WIDTH, &width, sizeof(width));
+        status |= vxQueryImage(input, VX_IMAGE_HEIGHT, &height, sizeof(height));
+        status |= vxQueryImage(input, VX_IMAGE_FORMAT, &format, sizeof(format));
+
+        status |= vxQueryPyramid(laplacian, VX_PYRAMID_LEVELS, &levels, sizeof(levels));
+
+        status |= vxQueryNode(node, VX_NODE_BORDER, &border, sizeof(border));
+        
+        border.mode = VX_BORDER_REPLICATE;
+
+        vxSetContextAttribute(context, VX_CONTEXT_IMMEDIATE_BORDER, &border, sizeof(border));
+
+        gaussian = vxCreatePyramid(context, levels + 1, VX_SCALE_PYRAMID_HALF, width, height, VX_DF_IMAGE_U8);
+        vxuGaussianPyramid(context, input, gaussian);
+
+        conv = vxCreateGaussian5x5Convolution(context);
+
+        level_width = width;
+        level_height = height;
+
+        gauss_cur = vxGetPyramidLevel(gaussian, 0);
+        gauss_next = vxGetPyramidLevel(gaussian, 1);
+        for (lev = 0; lev < levels; lev++)
+        {
+            pyr_gauss_curr_level_filtered = vxCreateImage(context, level_width, level_height, VX_DF_IMAGE_S16);
+            upsampleImage(context, level_width, level_height, gauss_next, conv, pyr_gauss_curr_level_filtered, &border);
+
+            pyr_laplacian_curr_level = vxGetPyramidLevel(laplacian, (vx_uint32)lev);
+            status |= vxuSubtract(context, gauss_cur, pyr_gauss_curr_level_filtered, policy, pyr_laplacian_curr_level);
+
+            if (lev == levels - 1)
+            {
+                vx_image tmp = vxGetPyramidLevel(gaussian, (vx_uint32) levels);
+                ownCopyImage(tmp, output);
+                vxReleaseImage(&tmp);
+                vxReleaseImage(&gauss_next);
+                vxReleaseImage(&gauss_cur);
+            }
+            else
+            {
+                /* compute dimensions for the next level */
+                level_width = (vx_uint32)ceilf(level_width * VX_SCALE_PYRAMID_HALF);
+                level_height = (vx_uint32)ceilf(level_height * VX_SCALE_PYRAMID_HALF);
+                /* prepare to the next iteration */
+                /* make the next level of gaussian pyramid the current level */
+                vxReleaseImage(&gauss_next);
+                vxReleaseImage(&gauss_cur);
+                gauss_cur = vxGetPyramidLevel(gaussian, (vx_uint32)lev + 1);
+                gauss_next = vxGetPyramidLevel(gaussian, (vx_uint32)lev + 2);
+
+            }
+
+            /* decrements the references */
+
+            status |= vxReleaseImage(&pyr_gauss_curr_level_filtered);
+            status |= vxReleaseImage(&pyr_laplacian_curr_level);
+        }
+
+        status |= vxReleasePyramid(&gaussian);
+        status |= vxReleaseConvolution(&conv);
     }
+
     return status;
+
 }
 
 static vx_status VX_CALLBACK vxLaplacianPyramidInputValidator(vx_node node, vx_uint32 index)
@@ -862,89 +942,8 @@ static vx_status VX_CALLBACK vxLaplacianPyramidInitializer(vx_node node, const v
 
     if (num == dimof(laplacian_pyramid_kernel_params))
     {
-        vx_context context = vxGetContext((vx_reference)node);
-
-        vx_size lev;
-        vx_size levels = 1;
-        vx_uint32 width = 0;
-        vx_uint32 height = 0;
-        vx_uint32 level_width = 0;
-        vx_uint32 level_height = 0;
-        vx_df_image format;
-        vx_enum policy = VX_CONVERT_POLICY_SATURATE;
-        vx_border_t border;
-        vx_convolution conv = 0;
-        vx_image pyr_gauss_curr_level_filtered = 0;
-        vx_image pyr_laplacian_curr_level = 0;
-        vx_image   input = (vx_image)parameters[0];
-        vx_pyramid laplacian = (vx_pyramid)parameters[1];
-        vx_image   output = (vx_image)parameters[2];
-        vx_pyramid gaussian = 0;
-        vx_image gauss_cur = 0;
-        vx_image gauss_next = 0;
-
-        status |= vxQueryImage(input, VX_IMAGE_WIDTH, &width, sizeof(width));
-        status |= vxQueryImage(input, VX_IMAGE_HEIGHT, &height, sizeof(height));
-        status |= vxQueryImage(input, VX_IMAGE_FORMAT, &format, sizeof(format));
-
-        status |= vxQueryPyramid(laplacian, VX_PYRAMID_LEVELS, &levels, sizeof(levels));
-
-        status |= vxQueryNode(node, VX_NODE_BORDER, &border, sizeof(border));
-        
-        border.mode = VX_BORDER_REPLICATE;
-
-        vxSetContextAttribute(context, VX_CONTEXT_IMMEDIATE_BORDER, &border, sizeof(border));
-
-        gaussian = vxCreatePyramid(context, levels + 1, VX_SCALE_PYRAMID_HALF, width, height, VX_DF_IMAGE_U8);
-        vxuGaussianPyramid(context, input, gaussian);
-
-        conv = vxCreateGaussian5x5Convolution(context);
-
-        level_width = width;
-        level_height = height;
-
-        gauss_cur = vxGetPyramidLevel(gaussian, 0);
-        gauss_next = vxGetPyramidLevel(gaussian, 1);
-        for (lev = 0; lev < levels; lev++)
-        {
-            pyr_gauss_curr_level_filtered = vxCreateImage(context, level_width, level_height, VX_DF_IMAGE_S16);
-            upsampleImage(context, level_width, level_height, gauss_next, conv, pyr_gauss_curr_level_filtered, &border);
-
-            pyr_laplacian_curr_level = vxGetPyramidLevel(laplacian, (vx_uint32)lev);
-            status |= vxuSubtract(context, gauss_cur, pyr_gauss_curr_level_filtered, policy, pyr_laplacian_curr_level);
-
-            if (lev == levels - 1)
-            {
-                vx_image tmp = vxGetPyramidLevel(gaussian, (vx_uint32) levels);
-                ownCopyImage(tmp, output);
-                vxReleaseImage(&tmp);
-                vxReleaseImage(&gauss_next);
-                vxReleaseImage(&gauss_cur);
-            }
-            else
-            {
-                /* compute dimensions for the next level */
-                level_width = (vx_uint32)ceilf(level_width * VX_SCALE_PYRAMID_HALF);
-                level_height = (vx_uint32)ceilf(level_height * VX_SCALE_PYRAMID_HALF);
-                /* prepare to the next iteration */
-                /* make the next level of gaussian pyramid the current level */
-                vxReleaseImage(&gauss_next);
-                vxReleaseImage(&gauss_cur);
-                gauss_cur = vxGetPyramidLevel(gaussian, (vx_uint32)lev + 1);
-                gauss_next = vxGetPyramidLevel(gaussian, (vx_uint32)lev + 2);
-
-            }
-
-            /* decrements the references */
-
-            status |= vxReleaseImage(&pyr_gauss_curr_level_filtered);
-            status |= vxReleaseImage(&pyr_laplacian_curr_level);
-        }
-
-        status |= vxReleasePyramid(&gaussian);
-        status |= vxReleaseConvolution(&conv);
+        /* computation moved to kernel */
     }
-
     return status;
 }
 
@@ -988,15 +987,88 @@ static vx_param_description_t laplacian_reconstruct_kernel_params[] =
 
 static vx_status VX_CALLBACK vxLaplacianReconstructKernel(vx_node node, const vx_reference parameters[], vx_uint32 num)
 {
-    vx_status status = VX_FAILURE;
-    (void)node;
-    (void)parameters;
+    vx_status status = VX_SUCCESS;
 
     if (num == dimof(laplacian_reconstruct_kernel_params))
     {
-        status = VX_SUCCESS;
+        vx_context context = vxGetContext((vx_reference)node);
+
+        vx_size lev;
+        vx_size levels = 1;
+        vx_uint32 width = 0;
+        vx_uint32 height = 0;
+        vx_uint32 level_width = 0;
+        vx_uint32 level_height = 0;
+        vx_df_image format = VX_DF_IMAGE_S16;
+        vx_enum policy = VX_CONVERT_POLICY_SATURATE;
+        vx_border_t border;
+        vx_image filling = 0;
+        vx_image pyr_level = 0;
+        vx_image filter = 0;
+        vx_image out = 0;
+        vx_convolution conv;
+
+        vx_pyramid laplacian = (vx_pyramid)parameters[0];
+        vx_image   input = (vx_image)parameters[1];
+        vx_image   output = (vx_image)parameters[2];
+
+        vx_scalar spolicy = vxCreateScalar(context, VX_TYPE_ENUM, &policy);
+
+        status |= vxQueryImage(input, VX_IMAGE_WIDTH, &width, sizeof(width));
+        status |= vxQueryImage(input, VX_IMAGE_HEIGHT, &height, sizeof(height));
+
+        status |= vxQueryPyramid(laplacian, VX_PYRAMID_LEVELS, &levels, sizeof(levels));
+
+        status |= vxQueryNode(node, VX_NODE_BORDER, &border, sizeof(border));
+        border.mode = VX_BORDER_REPLICATE;
+        conv = vxCreateGaussian5x5Convolution(context);
+
+        level_width = (vx_uint32)ceilf(width  * VX_SCALE_PYRAMID_DOUBLE);
+        level_height = (vx_uint32)ceilf(height * VX_SCALE_PYRAMID_DOUBLE);
+        filling = vxCreateImage(context, width, height, format);
+        for (lev = 0; lev < levels; lev++)
+        {
+            out = vxCreateImage(context, level_width, level_height, format);
+            filter = vxCreateImage(context, level_width, level_height, format);
+
+            pyr_level = vxGetPyramidLevel(laplacian, (vx_uint32)((levels - 1) - lev));
+
+            if (lev == 0)
+            {
+                ownCopyImage(input, filling);
+            }
+            upsampleImage(context, level_width, level_height, filling, conv, filter, &border);
+            vxAddition(filter, pyr_level, spolicy, out);
+
+            status |= vxReleaseImage(&pyr_level);
+
+            if ((levels - 1) - lev == 0)
+            {
+                ownCopyImage(out, output);
+                status |= vxReleaseImage(&filling);
+            }
+            else
+            {
+                /* compute dimensions for the next level */
+                status |= vxReleaseImage(&filling);
+                filling = vxCreateImage(context, level_width, level_height, format);
+                ownCopyImage(out, filling);
+
+                level_width = (vx_uint32)ceilf(level_width  * VX_SCALE_PYRAMID_DOUBLE);
+                level_height = (vx_uint32)ceilf(level_height * VX_SCALE_PYRAMID_DOUBLE);
+
+
+            }
+            status |= vxReleaseImage(&out);
+            status |= vxReleaseImage(&filter);
+
+        }
+        status |= vxReleaseConvolution(&conv);
+        status |= vxReleaseScalar(&spolicy);
     }
+
     return status;
+
 }
 
 static vx_status VX_CALLBACK vxLaplacianReconstructInputValidator(vx_node node, vx_uint32 index)
@@ -1144,82 +1216,8 @@ static vx_status VX_CALLBACK vxLaplacianReconstructInitializer(vx_node node, con
 
     if (num == dimof(laplacian_reconstruct_kernel_params))
     {
-        vx_context context = vxGetContext((vx_reference)node);
-
-        vx_size lev;
-        vx_size levels = 1;
-        vx_uint32 width = 0;
-        vx_uint32 height = 0;
-        vx_uint32 level_width = 0;
-        vx_uint32 level_height = 0;
-        vx_df_image format = VX_DF_IMAGE_S16;
-        vx_enum policy = VX_CONVERT_POLICY_SATURATE;
-        vx_border_t border;
-        vx_image filling = 0;
-        vx_image pyr_level = 0;
-        vx_image filter = 0;
-        vx_image out = 0;
-        vx_convolution conv;
-
-        vx_pyramid laplacian = (vx_pyramid)parameters[0];
-        vx_image   input = (vx_image)parameters[1];
-        vx_image   output = (vx_image)parameters[2];
-
-        vx_scalar spolicy = vxCreateScalar(context, VX_TYPE_ENUM, &policy);
-
-        status |= vxQueryImage(input, VX_IMAGE_WIDTH, &width, sizeof(width));
-        status |= vxQueryImage(input, VX_IMAGE_HEIGHT, &height, sizeof(height));
-
-        status |= vxQueryPyramid(laplacian, VX_PYRAMID_LEVELS, &levels, sizeof(levels));
-
-        status |= vxQueryNode(node, VX_NODE_BORDER, &border, sizeof(border));
-        border.mode = VX_BORDER_REPLICATE;
-        conv = vxCreateGaussian5x5Convolution(context);
-
-        level_width = (vx_uint32)ceilf(width  * VX_SCALE_PYRAMID_DOUBLE);
-        level_height = (vx_uint32)ceilf(height * VX_SCALE_PYRAMID_DOUBLE);
-        filling = vxCreateImage(context, width, height, format);
-        for (lev = 0; lev < levels; lev++)
-        {
-            out = vxCreateImage(context, level_width, level_height, format);
-            filter = vxCreateImage(context, level_width, level_height, format);
-
-            pyr_level = vxGetPyramidLevel(laplacian, (vx_uint32)((levels - 1) - lev));
-
-            if (lev == 0)
-            {
-                ownCopyImage(input, filling);
-            }
-            upsampleImage(context, level_width, level_height, filling, conv, filter, &border);
-            vxAddition(filter, pyr_level, spolicy, out);
-
-            status |= vxReleaseImage(&pyr_level);
-
-            if ((levels - 1) - lev == 0)
-            {
-                ownCopyImage(out, output);
-                status |= vxReleaseImage(&filling);
-            }
-            else
-            {
-                /* compute dimensions for the next level */
-                status |= vxReleaseImage(&filling);
-                filling = vxCreateImage(context, level_width, level_height, format);
-                ownCopyImage(out, filling);
-
-                level_width = (vx_uint32)ceilf(level_width  * VX_SCALE_PYRAMID_DOUBLE);
-                level_height = (vx_uint32)ceilf(level_height * VX_SCALE_PYRAMID_DOUBLE);
-
-
-            }
-            status |= vxReleaseImage(&out);
-            status |= vxReleaseImage(&filter);
-
-        }
-        status |= vxReleaseConvolution(&conv);
-        status |= vxReleaseScalar(&spolicy);
+        /* computation moved to kernel */
     }
-
     return status;
 }
 
diff --git a/sample/targets/venum/vx_laplacianpyramid.c b/sample/targets/venum/vx_laplacianpyramid.c
index 42c3fc2..4c80331 100644
--- a/sample/targets/venum/vx_laplacianpyramid.c
+++ b/sample/targets/venum/vx_laplacianpyramid.c
@@ -229,12 +229,93 @@ static vx_param_description_t laplacian_pyramid_kernel_params[] =
 static vx_status VX_CALLBACK vxLaplacianPyramidKernel(vx_node node, const vx_reference parameters[], vx_uint32 num)
 
 {
+    vx_status status = VX_SUCCESS;
 
-    vx_status status = VX_FAILURE;
     if (num == dimof(laplacian_pyramid_kernel_params))
     {
-        status = VX_SUCCESS;
+        vx_context context = vxGetContext((vx_reference)node);
+
+        vx_size lev;
+        vx_size levels = 1;
+        vx_uint32 width = 0;
+        vx_uint32 height = 0;
+        vx_uint32 level_width = 0;
+        vx_uint32 level_height = 0;
+        vx_df_image format;
+        vx_enum policy = VX_CONVERT_POLICY_SATURATE;
+        vx_border_t border;
+        vx_convolution conv = 0;
+        vx_image pyr_gauss_curr_level_filtered = 0;
+        vx_image pyr_laplacian_curr_level = 0;
+        vx_image   input = (vx_image)parameters[0];
+        vx_pyramid laplacian = (vx_pyramid)parameters[1];
+        vx_image   output = (vx_image)parameters[2];
+        vx_pyramid gaussian = 0;
+        vx_image gauss_cur = 0;
+        vx_image gauss_next = 0;
+        vx_image upsample = 0;
+
+        status |= vxQueryImage(input, VX_IMAGE_WIDTH, &width, sizeof(width));
+        status |= vxQueryImage(input, VX_IMAGE_HEIGHT, &height, sizeof(height));
+        status |= vxQueryImage(input, VX_IMAGE_FORMAT, &format, sizeof(format));
+
+        status |= vxQueryPyramid(laplacian, VX_PYRAMID_LEVELS, &levels, sizeof(levels));
+
+        status |= vxQueryNode(node, VX_NODE_BORDER, &border, sizeof(border));
+        
+        border.mode = VX_BORDER_REPLICATE;
+
+        gaussian = vxCreatePyramid(context, levels + 1, VX_SCALE_PYRAMID_HALF, width, height, VX_DF_IMAGE_U8);
+        vxuGaussianPyramid(context, input, gaussian);
+
+        conv = vxCreateGaussian5x5Convolution(context);
+
+        level_width = width;
+        level_height = height;
+
+        gauss_cur = vxGetPyramidLevel(gaussian, 0);
+        gauss_next = vxGetPyramidLevel(gaussian, 1);
+        for (lev = 0; lev < levels; lev++)
+        {
+            pyr_gauss_curr_level_filtered = vxCreateImage(context, level_width, level_height, VX_DF_IMAGE_S16);
+            xLaplacianPyramidupsampleImage(context, level_width, level_height, gauss_next, conv, pyr_gauss_curr_level_filtered, &border);
+
+            pyr_laplacian_curr_level = vxGetPyramidLevel(laplacian, (vx_uint32)lev);
+            status |= vxuSubtract(context, gauss_cur, pyr_gauss_curr_level_filtered, policy, pyr_laplacian_curr_level);
+            vxReleaseImage(&upsample);
+
+            if (lev == levels - 1)
+            {
+                vx_image tmp = vxGetPyramidLevel(gaussian, levels);
+                ownCopyImage(tmp, output);
+                vxReleaseImage(&tmp);
+                vxReleaseImage(&gauss_next);
+                vxReleaseImage(&gauss_cur);
+            }
+            else
+            {
+                /* compute dimensions for the next level */
+                level_width = (vx_uint32)ceilf(level_width * VX_SCALE_PYRAMID_HALF);
+                level_height = (vx_uint32)ceilf(level_height * VX_SCALE_PYRAMID_HALF);
+                /* prepare to the next iteration */
+                /* make the next level of gaussian pyramid the current level */
+                vxReleaseImage(&gauss_next);
+                vxReleaseImage(&gauss_cur);
+                gauss_cur = vxGetPyramidLevel(gaussian, lev + 1);
+                gauss_next = vxGetPyramidLevel(gaussian, lev + 2);
+
+            }
+
+            /* decrements the references */
+
+            status |= vxReleaseImage(&pyr_gauss_curr_level_filtered);
+            status |= vxReleaseImage(&pyr_laplacian_curr_level);
+        }
+
+        status |= vxReleasePyramid(&gaussian);
+        status |= vxReleaseConvolution(&conv);
     }
+
     return status;
 
 }
@@ -398,89 +479,8 @@ static vx_status VX_CALLBACK vxLaplacianPyramidInitializer(vx_node node, const v
 
     if (num == dimof(laplacian_pyramid_kernel_params))
     {
-        vx_context context = vxGetContext((vx_reference)node);
-
-        vx_size lev;
-        vx_size levels = 1;
-        vx_uint32 width = 0;
-        vx_uint32 height = 0;
-        vx_uint32 level_width = 0;
-        vx_uint32 level_height = 0;
-        vx_df_image format;
-        vx_enum policy = VX_CONVERT_POLICY_SATURATE;
-        vx_border_t border;
-        vx_convolution conv = 0;
-        vx_image pyr_gauss_curr_level_filtered = 0;
-        vx_image pyr_laplacian_curr_level = 0;
-        vx_image   input = (vx_image)parameters[0];
-        vx_pyramid laplacian = (vx_pyramid)parameters[1];
-        vx_image   output = (vx_image)parameters[2];
-        vx_pyramid gaussian = 0;
-        vx_image gauss_cur = 0;
-        vx_image gauss_next = 0;
-        vx_image upsample = 0;
-
-        status |= vxQueryImage(input, VX_IMAGE_WIDTH, &width, sizeof(width));
-        status |= vxQueryImage(input, VX_IMAGE_HEIGHT, &height, sizeof(height));
-        status |= vxQueryImage(input, VX_IMAGE_FORMAT, &format, sizeof(format));
-
-        status |= vxQueryPyramid(laplacian, VX_PYRAMID_LEVELS, &levels, sizeof(levels));
-
-        status |= vxQueryNode(node, VX_NODE_BORDER, &border, sizeof(border));
-        
-        border.mode = VX_BORDER_REPLICATE;
-
-        gaussian = vxCreatePyramid(context, levels + 1, VX_SCALE_PYRAMID_HALF, width, height, VX_DF_IMAGE_U8);
-        vxuGaussianPyramid(context, input, gaussian);
-
-        conv = vxCreateGaussian5x5Convolution(context);
-
-        level_width = width;
-        level_height = height;
-
-        gauss_cur = vxGetPyramidLevel(gaussian, 0);
-        gauss_next = vxGetPyramidLevel(gaussian, 1);
-        for (lev = 0; lev < levels; lev++)
-        {
-            pyr_gauss_curr_level_filtered = vxCreateImage(context, level_width, level_height, VX_DF_IMAGE_S16);
-            xLaplacianPyramidupsampleImage(context, level_width, level_height, gauss_next, conv, pyr_gauss_curr_level_filtered, &border);
-
-            pyr_laplacian_curr_level = vxGetPyramidLevel(laplacian, (vx_uint32)lev);
-            status |= vxuSubtract(context, gauss_cur, pyr_gauss_curr_level_filtered, policy, pyr_laplacian_curr_level);
-            vxReleaseImage(&upsample);
-
-            if (lev == levels - 1)
-            {
-                vx_image tmp = vxGetPyramidLevel(gaussian, levels);
-                ownCopyImage(tmp, output);
-                vxReleaseImage(&tmp);
-                vxReleaseImage(&gauss_next);
-                vxReleaseImage(&gauss_cur);
-            }
-            else
-            {
-                /* compute dimensions for the next level */
-                level_width = (vx_uint32)ceilf(level_width * VX_SCALE_PYRAMID_HALF);
-                level_height = (vx_uint32)ceilf(level_height * VX_SCALE_PYRAMID_HALF);
-                /* prepare to the next iteration */
-                /* make the next level of gaussian pyramid the current level */
-                vxReleaseImage(&gauss_next);
-                vxReleaseImage(&gauss_cur);
-                gauss_cur = vxGetPyramidLevel(gaussian, lev + 1);
-                gauss_next = vxGetPyramidLevel(gaussian, lev + 2);
-
-            }
-
-            /* decrements the references */
-
-            status |= vxReleaseImage(&pyr_gauss_curr_level_filtered);
-            status |= vxReleaseImage(&pyr_laplacian_curr_level);
-        }
-
-        status |= vxReleasePyramid(&gaussian);
-        status |= vxReleaseConvolution(&conv);
+        /* computation moved to kernel */
     }
-
     return status;
 }
 
diff --git a/sample/targets/venum/vx_pyramid.c b/sample/targets/venum/vx_pyramid.c
index 5959f55..9f45956 100644
--- a/sample/targets/venum/vx_pyramid.c
+++ b/sample/targets/venum/vx_pyramid.c
@@ -741,12 +741,88 @@ static vx_param_description_t laplacian_reconstruct_kernel_params[] =
 
 static vx_status VX_CALLBACK vxLaplacianReconstructKernel(vx_node node, const vx_reference parameters[], vx_uint32 num)
 {
-    vx_status status = VX_FAILURE;
+    vx_status status = VX_SUCCESS;
+
     if (num == dimof(laplacian_reconstruct_kernel_params))
     {
-        status = VX_SUCCESS;
+        vx_context context = vxGetContext((vx_reference)node);
+
+        vx_size lev;
+        vx_size levels = 1;
+        vx_uint32 width = 0;
+        vx_uint32 height = 0;
+        vx_uint32 level_width = 0;
+        vx_uint32 level_height = 0;
+        vx_df_image format = VX_DF_IMAGE_S16;
+        vx_enum policy = VX_CONVERT_POLICY_SATURATE;
+        vx_border_t border;
+        vx_image filling = 0;
+        vx_image pyr_level = 0;
+        vx_image filter = 0;
+        vx_image out = 0;
+        vx_convolution conv;
+
+        vx_pyramid laplacian = (vx_pyramid)parameters[0];
+        vx_image   input = (vx_image)parameters[1];
+        vx_image   output = (vx_image)parameters[2];
+
+        vx_scalar spolicy = vxCreateScalar(context, VX_TYPE_ENUM, &policy);
+
+        status |= vxQueryImage(input, VX_IMAGE_WIDTH, &width, sizeof(width));
+        status |= vxQueryImage(input, VX_IMAGE_HEIGHT, &height, sizeof(height));
+
+        status |= vxQueryPyramid(laplacian, VX_PYRAMID_LEVELS, &levels, sizeof(levels));
+
+        status |= vxQueryNode(node, VX_NODE_BORDER, &border, sizeof(border));
+        border.mode = VX_BORDER_REPLICATE;
+        conv = vxCreateGaussian5x5Convolution(context);
+
+        level_width = (vx_uint32)ceilf(width  * VX_SCALE_PYRAMID_DOUBLE);
+        level_height = (vx_uint32)ceilf(height * VX_SCALE_PYRAMID_DOUBLE);
+        filling = vxCreateImage(context, width, height, format);
+        for (lev = 0; lev < levels; lev++)
+        {
+            out = vxCreateImage(context, level_width, level_height, format);
+            filter = vxCreateImage(context, level_width, level_height, format);
+
+            pyr_level = vxGetPyramidLevel(laplacian, (vx_uint32)((levels - 1) - lev));
+
+            if (lev == 0)
+            {
+                ownCopyImage_S16(input, filling);
+            }
+            upsampleImage(context, level_width, level_height, filling, conv, filter, &border);
+            vxAddition(filter, pyr_level, spolicy, out);
+
+            status |= vxReleaseImage(&pyr_level);
+
+            if ((levels - 1) - lev == 0)
+            {
+                ownCopyImage(out, output);
+                status |= vxReleaseImage(&filling);
+            }
+            else
+            {
+                /* compute dimensions for the next level */
+                status |= vxReleaseImage(&filling);
+                filling = vxCreateImage(context, level_width, level_height, format);
+                ownCopyImage(out, filling);
+
+                level_width = (vx_uint32)ceilf(level_width  * VX_SCALE_PYRAMID_DOUBLE);
+                level_height = (vx_uint32)ceilf(level_height * VX_SCALE_PYRAMID_DOUBLE);
+
+
+            }
+            status |= vxReleaseImage(&out);
+            status |= vxReleaseImage(&filter);
+
+        }
+        status |= vxReleaseConvolution(&conv);
+        status |= vxReleaseScalar(&spolicy);
     }
+
     return status;
+
 }
 
 static vx_status VX_CALLBACK vxLaplacianReconstructInputValidator(vx_node node, vx_uint32 index)
@@ -967,82 +1043,8 @@ static vx_status VX_CALLBACK vxLaplacianReconstructInitializer(vx_node node, con
 
     if (num == dimof(laplacian_reconstruct_kernel_params))
     {
-        vx_context context = vxGetContext((vx_reference)node);
-
-        vx_size lev;
-        vx_size levels = 1;
-        vx_uint32 width = 0;
-        vx_uint32 height = 0;
-        vx_uint32 level_width = 0;
-        vx_uint32 level_height = 0;
-        vx_df_image format = VX_DF_IMAGE_S16;
-        vx_enum policy = VX_CONVERT_POLICY_SATURATE;
-        vx_border_t border;
-        vx_image filling = 0;
-        vx_image pyr_level = 0;
-        vx_image filter = 0;
-        vx_image out = 0;
-        vx_convolution conv;
-
-        vx_pyramid laplacian = (vx_pyramid)parameters[0];
-        vx_image   input = (vx_image)parameters[1];
-        vx_image   output = (vx_image)parameters[2];
-
-        vx_scalar spolicy = vxCreateScalar(context, VX_TYPE_ENUM, &policy);
-
-        status |= vxQueryImage(input, VX_IMAGE_WIDTH, &width, sizeof(width));
-        status |= vxQueryImage(input, VX_IMAGE_HEIGHT, &height, sizeof(height));
-
-        status |= vxQueryPyramid(laplacian, VX_PYRAMID_LEVELS, &levels, sizeof(levels));
-
-        status |= vxQueryNode(node, VX_NODE_BORDER, &border, sizeof(border));
-        border.mode = VX_BORDER_REPLICATE;
-        conv = vxCreateGaussian5x5Convolution(context);
-
-        level_width = (vx_uint32)ceilf(width  * VX_SCALE_PYRAMID_DOUBLE);
-        level_height = (vx_uint32)ceilf(height * VX_SCALE_PYRAMID_DOUBLE);
-        filling = vxCreateImage(context, width, height, format);
-        for (lev = 0; lev < levels; lev++)
-        {
-            out = vxCreateImage(context, level_width, level_height, format);
-            filter = vxCreateImage(context, level_width, level_height, format);
-
-            pyr_level = vxGetPyramidLevel(laplacian, (vx_uint32)((levels - 1) - lev));
-
-            if (lev == 0)
-            {
-                ownCopyImage_S16(input, filling);
-            }
-            upsampleImage(context, level_width, level_height, filling, conv, filter, &border);
-            vxAddition(filter, pyr_level, spolicy, out);
-
-            status |= vxReleaseImage(&pyr_level);
-
-            if ((levels - 1) - lev == 0)
-            {
-                ownCopyImage(out, output);
-                status |= vxReleaseImage(&filling);
-            }
-            else
-            {
-                /* compute dimensions for the next level */
-                status |= vxReleaseImage(&filling);
-                filling = vxCreateImage(context, level_width, level_height, format);
-                ownCopyImage(out, filling);
-
-                level_width = (vx_uint32)ceilf(level_width  * VX_SCALE_PYRAMID_DOUBLE);
-                level_height = (vx_uint32)ceilf(level_height * VX_SCALE_PYRAMID_DOUBLE);
-
-
-            }
-            status |= vxReleaseImage(&out);
-            status |= vxReleaseImage(&filter);
-
-        }
-        status |= vxReleaseConvolution(&conv);
-        status |= vxReleaseScalar(&spolicy);
+        /* computation moved to kernel */
     }
-
     return status;
 }