COMPMID-534: Port MemoryManager to CL functions (Images)

Adds support for:
-CLCannyEdge
-CLFastCorners
-CLGaussian5x5
-CLHarrisCorners
-CLSobel3x3
-CLSobel5x5

Change-Id: I712a76d4ceda915b5cf85a4d12c1b7a059d4d909
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/88118
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

13 files changed, 134 insertions, 30 deletions

diff --git a/arm_compute/runtime/CL/functions/CLCannyEdge.h b/arm_compute/runtime/CL/functions/CLCannyEdge.h
index e5a82b2263..1d5a5aaeaa 100644
--- a/arm_compute/runtime/CL/functions/CLCannyEdge.h
+++ b/arm_compute/runtime/CL/functions/CLCannyEdge.h
@@ -28,7 +28,9 @@
 
 #include "arm_compute/core/CL/kernels/CLCannyEdgeKernel.h"
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/IMemoryManager.h"
 
 #include <memory>
 
@@ -49,7 +51,7 @@ class CLCannyEdge : public IFunction
 {
 public:
     /** Constructor */
-    CLCannyEdge();
+    CLCannyEdge(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Initialise the function's source, destination, thresholds, gradient size, normalization type and border mode.
      *
      * @param[in,out] input                 Source tensor. Data types supported: U8. (Written to only for border_mode != UNDEFINED)
@@ -68,6 +70,7 @@ public:
     virtual void run() override;
 
 private:
+    CLMemoryGroup                 _memory_group;                                    /**< Function's memory group */
     std::unique_ptr<IFunction>    _sobel;                                           /**< Pointer to Sobel kernel. */
     CLGradientKernel              _gradient;                                        /**< Gradient kernel. */
     CLFillBorderKernel            _border_mag_gradient;                             /**< Fill border on magnitude tensor kernel */
diff --git a/arm_compute/runtime/CL/functions/CLFastCorners.h b/arm_compute/runtime/CL/functions/CLFastCorners.h
index 79d82af462..9afec71bc3 100644
--- a/arm_compute/runtime/CL/functions/CLFastCorners.h
+++ b/arm_compute/runtime/CL/functions/CLFastCorners.h
@@ -29,11 +29,14 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Window.h"
 #include "arm_compute/runtime/CL/CLArray.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/CL/functions/CLNonMaximaSuppression3x3.h"
 #include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
 
 #include <cstdint>
+#include <memory>
 
 namespace arm_compute
 {
@@ -51,7 +54,7 @@ class CLFastCorners : public IFunction
 {
 public:
     /** Constructor */
-    CLFastCorners();
+    CLFastCorners(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     CLFastCorners(const CLFastCorners &) = delete;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
@@ -72,6 +75,7 @@ public:
     void run() override;
 
 private:
+    CLMemoryGroup             _memory_group;
     CLFastCornersKernel       _fast_corners_kernel;
     CLNonMaximaSuppression3x3 _suppr_func;
     CLCopyToArrayKernel       _copy_array_kernel;
diff --git a/arm_compute/runtime/CL/functions/CLGaussian5x5.h b/arm_compute/runtime/CL/functions/CLGaussian5x5.h
index 148b9a9924..3c60cc66a3 100644
--- a/arm_compute/runtime/CL/functions/CLGaussian5x5.h
+++ b/arm_compute/runtime/CL/functions/CLGaussian5x5.h
@@ -27,10 +27,13 @@
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
 #include "arm_compute/core/CL/kernels/CLGaussian5x5Kernel.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
 
 #include <cstdint>
+#include <memory>
 
 namespace arm_compute
 {
@@ -47,7 +50,7 @@ class CLGaussian5x5 : public IFunction
 {
 public:
     /** Default Constructor. */
-    CLGaussian5x5();
+    CLGaussian5x5(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Initialise the function's source, destinations and border mode.
      *
      * @param[in,out] input                 Source tensor. Data types supported: U8. (Written to only for @p border_mode != UNDEFINED)
@@ -61,6 +64,7 @@ public:
     void run() override;
 
 protected:
+    CLMemoryGroup           _memory_group;   /**< Function's memory group */
     CLGaussian5x5HorKernel  _kernel_hor;     /**< Horizontal pass kernel */
     CLGaussian5x5VertKernel _kernel_vert;    /**< Vertical pass kernel */
     CLFillBorderKernel      _border_handler; /**< Kernel to handle image borders */
diff --git a/arm_compute/runtime/CL/functions/CLHarrisCorners.h b/arm_compute/runtime/CL/functions/CLHarrisCorners.h
index f9a1275f68..e09e67060f 100644
--- a/arm_compute/runtime/CL/functions/CLHarrisCorners.h
+++ b/arm_compute/runtime/CL/functions/CLHarrisCorners.h
@@ -31,11 +31,12 @@
 #include "arm_compute/core/CL/kernels/CLHarrisCornersKernel.h"
 #include "arm_compute/core/NEON/kernels/NEHarrisCornersKernel.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/CL/functions/CLNonMaximaSuppression3x3.h"
+#include "arm_compute/runtime/IMemoryManager.h"
 
 #include <cstdint>
-
 #include <memory>
 
 namespace arm_compute
@@ -60,7 +61,7 @@ class CLHarrisCorners : public IFunction
 {
 public:
     /** Constructor */
-    CLHarrisCorners();
+    CLHarrisCorners(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     CLHarrisCorners(const CLHarrisCorners &) = delete;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
@@ -85,6 +86,7 @@ public:
     void run() override;
 
 private:
+    CLMemoryGroup                       _memory_group;          /**< Function's memory group */
     std::unique_ptr<IFunction>          _sobel;                 /**< Sobel function */
     CLHarrisScoreKernel                 _harris_score;          /**< Harris score kernel */
     CLNonMaximaSuppression3x3           _non_max_suppr;         /**< Non-maxima suppression function */
diff --git a/arm_compute/runtime/CL/functions/CLSobel5x5.h b/arm_compute/runtime/CL/functions/CLSobel5x5.h
index ad1f72faf8..3e603f8311 100644
--- a/arm_compute/runtime/CL/functions/CLSobel5x5.h
+++ b/arm_compute/runtime/CL/functions/CLSobel5x5.h
@@ -27,10 +27,13 @@
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
 #include "arm_compute/core/CL/kernels/CLSobel5x5Kernel.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
 
 #include <cstdint>
+#include <memory>
 
 namespace arm_compute
 {
@@ -47,7 +50,7 @@ class CLSobel5x5 : public IFunction
 {
 public:
     /** Default Constructor. */
-    CLSobel5x5();
+    CLSobel5x5(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Initialise the function's source, destinations and border mode.
      *
      * @note At least one of output_x or output_y must be not NULL.
@@ -64,6 +67,7 @@ public:
     void run() override;
 
 protected:
+    CLMemoryGroup        _memory_group;   /**< Function's memory group */
     CLSobel5x5HorKernel  _sobel_hor;      /**< Sobel Horizontal 5x5 kernel */
     CLSobel5x5VertKernel _sobel_vert;     /**< Sobel Vertical 5x5 kernel */
     CLFillBorderKernel   _border_handler; /**< Kernel to handle image borders */
diff --git a/arm_compute/runtime/CL/functions/CLSobel7x7.h b/arm_compute/runtime/CL/functions/CLSobel7x7.h
index 1a3fe1a50a..0dc0a1c5e9 100644
--- a/arm_compute/runtime/CL/functions/CLSobel7x7.h
+++ b/arm_compute/runtime/CL/functions/CLSobel7x7.h
@@ -27,10 +27,13 @@
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
 #include "arm_compute/core/CL/kernels/CLSobel7x7Kernel.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
 
 #include <cstdint>
+#include <memory>
 
 namespace arm_compute
 {
@@ -47,7 +50,7 @@ class CLSobel7x7 : public IFunction
 {
 public:
     /** Default Constructor. */
-    CLSobel7x7();
+    CLSobel7x7(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Initialise the function's source, destinations and border mode.
      *
      * @note At least one of output_x or output_y must be not NULL.
@@ -64,6 +67,7 @@ public:
     void run() override;
 
 protected:
+    CLMemoryGroup        _memory_group;   /**< Function's memory group */
     CLSobel7x7HorKernel  _sobel_hor;      /**< Sobel Horizontal 7x7 kernel */
     CLSobel7x7VertKernel _sobel_vert;     /**< Sobel Vertical 7x7 kernel */
     CLFillBorderKernel   _border_handler; /**< Kernel to handle image borders */
diff --git a/src/runtime/CL/functions/CLCannyEdge.cpp b/src/runtime/CL/functions/CLCannyEdge.cpp
index 448ca9289d..5acb8e7ddb 100644
--- a/src/runtime/CL/functions/CLCannyEdge.cpp
+++ b/src/runtime/CL/functions/CLCannyEdge.cpp
@@ -35,8 +35,9 @@
 
 using namespace arm_compute;
 
-CLCannyEdge::CLCannyEdge() // NOLINT
-    : _sobel(),
+CLCannyEdge::CLCannyEdge(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
+    : _memory_group(std::move(memory_manager)),
+      _sobel(),
       _gradient(),
       _border_mag_gradient(),
       _non_max_suppr(),
@@ -96,6 +97,10 @@ void CLCannyEdge::configure(ICLTensor *input, ICLTensor *output, int32_t upper_t
     TensorInfo info_s32(shape_l1_stack, 1, arm_compute::DataType::S32);
     _l1_stack.allocator()->init(info_s32);
 
+    // Manage intermediate buffers
+    _memory_group.manage(&_gx);
+    _memory_group.manage(&_gy);
+
     // Configure/Init sobelNxN
     if(gradient_size == 3)
     {
@@ -120,23 +125,43 @@ void CLCannyEdge::configure(ICLTensor *input, ICLTensor *output, int32_t upper_t
         ARM_COMPUTE_ERROR("Gradient %d size not supported", gradient_size);
     }
 
+    // Manage intermediate buffers
+    _memory_group.manage(&_mag);
+    _memory_group.manage(&_phase);
+
     // Configure gradient
     _gradient.configure(&_gx, &_gy, &_mag, &_phase, norm_type);
 
+    // Allocate intermediate buffers
+    _gx.allocator()->allocate();
+    _gy.allocator()->allocate();
+
+    // Manage intermediate buffers
+    _memory_group.manage(&_nonmax);
+
     // Configure non-maxima suppression
     _non_max_suppr.configure(&_mag, &_phase, &_nonmax, lower_thr, border_mode == BorderMode::UNDEFINED);
 
+    // Allocate intermediate buffers
+    _phase.allocator()->allocate();
+
     // Fill border around magnitude image as non-maxima suppression will access
     // it. If border mode is undefined filling the border is a nop.
     _border_mag_gradient.configure(&_mag, _non_max_suppr.border_size(), border_mode, constant_border_value);
 
+    // Allocate intermediate buffers
+    _mag.allocator()->allocate();
+
+    // Manage intermediate buffers
+    _memory_group.manage(&_visited);
+    _memory_group.manage(&_recorded);
+    _memory_group.manage(&_l1_stack);
+    _memory_group.manage(&_l1_list_counter);
+
     // Configure edge tracing
     _edge_trace.configure(&_nonmax, output, upper_thr, lower_thr, &_visited, &_recorded, &_l1_stack, &_l1_list_counter);
 
-    _gx.allocator()->allocate();
-    _gy.allocator()->allocate();
-    _phase.allocator()->allocate();
-    _mag.allocator()->allocate();
+    // Allocate intermediate buffers
     _visited.allocator()->allocate();
     _recorded.allocator()->allocate();
     _l1_stack.allocator()->allocate();
@@ -146,6 +171,8 @@ void CLCannyEdge::configure(ICLTensor *input, ICLTensor *output, int32_t upper_t
 
 void CLCannyEdge::run()
 {
+    _memory_group.acquire();
+
     // Run sobel
     _sobel->run();
 
@@ -165,4 +192,6 @@ void CLCannyEdge::run()
     _l1_list_counter.clear(CLScheduler::get().queue());
     _l1_stack.clear(CLScheduler::get().queue());
     CLScheduler::get().enqueue(_edge_trace, true);
+
+    _memory_group.release();
 }
diff --git a/src/runtime/CL/functions/CLFastCorners.cpp b/src/runtime/CL/functions/CLFastCorners.cpp
index d2903fb849..7a0dd09fbe 100644
--- a/src/runtime/CL/functions/CLFastCorners.cpp
+++ b/src/runtime/CL/functions/CLFastCorners.cpp
@@ -36,8 +36,9 @@
 
 using namespace arm_compute;
 
-CLFastCorners::CLFastCorners()
-    : _fast_corners_kernel(),
+CLFastCorners::CLFastCorners(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)),
+      _fast_corners_kernel(),
       _suppr_func(),
       _copy_array_kernel(),
       _output(),
@@ -70,6 +71,7 @@ void CLFastCorners::configure(const ICLImage *input, float threshold, bool nonma
 
     const bool update_number = (nullptr != _num_corners);
 
+    _memory_group.manage(&_output);
     _fast_corners_kernel.configure(input, &_output, threshold, nonmax_suppression, border_mode);
 
     if(!_non_max)
@@ -79,6 +81,7 @@ void CLFastCorners::configure(const ICLImage *input, float threshold, bool nonma
     else
     {
         _suppr.allocator()->init(tensor_info);
+        _memory_group.manage(&_suppr);
 
         _suppr_func.configure(&_output, &_suppr, border_mode);
         _copy_array_kernel.configure(&_suppr, update_number, corners, &_num_buffer);
@@ -94,6 +97,8 @@ void CLFastCorners::run()
 {
     cl::CommandQueue q = CLScheduler::get().queue();
 
+    _memory_group.acquire();
+
     if(_non_max)
     {
         ARM_COMPUTE_ERROR_ON_MSG(_output.cl_buffer().get() == nullptr, "Unconfigured function");
@@ -124,4 +129,6 @@ void CLFastCorners::run()
     }
 
     q.flush();
+
+    _memory_group.release();
 }
diff --git a/src/runtime/CL/functions/CLGaussian5x5.cpp b/src/runtime/CL/functions/CLGaussian5x5.cpp
index e83a8fb857..f30eee1df7 100644
--- a/src/runtime/CL/functions/CLGaussian5x5.cpp
+++ b/src/runtime/CL/functions/CLGaussian5x5.cpp
@@ -35,8 +35,8 @@
 
 using namespace arm_compute;
 
-CLGaussian5x5::CLGaussian5x5()
-    : _kernel_hor(), _kernel_vert(), _border_handler(), _tmp()
+CLGaussian5x5::CLGaussian5x5(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _kernel_hor(), _kernel_vert(), _border_handler(), _tmp()
 {
 }
 
@@ -46,6 +46,10 @@ void CLGaussian5x5::configure(ICLTensor *input, ICLTensor *output, BorderMode bo
 
     _tmp.allocator()->init(TensorInfo(input->info()->tensor_shape(), 1, DataType::U16));
 
+    // Manage intermediate buffers
+    _memory_group.manage(&_tmp);
+
+    // Configure kernels
     _kernel_hor.configure(input, &_tmp, border_mode == BorderMode::UNDEFINED);
     _kernel_vert.configure(&_tmp, output, border_mode == BorderMode::UNDEFINED);
     _border_handler.configure(input, _kernel_hor.border_size(), border_mode, PixelValue(constant_border_value));
@@ -57,6 +61,11 @@ void CLGaussian5x5::configure(ICLTensor *input, ICLTensor *output, BorderMode bo
 void CLGaussian5x5::run()
 {
     CLScheduler::get().enqueue(_border_handler, false);
+
+    _memory_group.acquire();
+
     CLScheduler::get().enqueue(_kernel_hor, false);
     CLScheduler::get().enqueue(_kernel_vert);
+
+    _memory_group.release();
 }
diff --git a/src/runtime/CL/functions/CLHarrisCorners.cpp b/src/runtime/CL/functions/CLHarrisCorners.cpp
index 2140240753..059528fe30 100644
--- a/src/runtime/CL/functions/CLHarrisCorners.cpp
+++ b/src/runtime/CL/functions/CLHarrisCorners.cpp
@@ -42,8 +42,9 @@
 
 using namespace arm_compute;
 
-CLHarrisCorners::CLHarrisCorners() // NOLINT
-    : _sobel(nullptr),
+CLHarrisCorners::CLHarrisCorners(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
+    : _memory_group(std::move(memory_manager)),
+      _sobel(nullptr),
       _harris_score(),
       _non_max_suppr(),
       _candidates(),
@@ -84,6 +85,10 @@ void CLHarrisCorners::configure(ICLImage *input, float threshold, float min_dist
 
     _corners_list = arm_compute::support::cpp14::make_unique<InternalKeypoint[]>(shape.x() * shape.y());
 
+    // Manage intermediate buffers
+    _memory_group.manage(&_gx);
+    _memory_group.manage(&_gy);
+
     /* Set/init Sobel kernel accordingly with gradient_size */
     switch(gradient_size)
     {
@@ -116,6 +121,9 @@ void CLHarrisCorners::configure(ICLImage *input, float threshold, float min_dist
     const float norm_factor               = 1.0f / (255.0f * pow(4.0f, gradient_size / 2) * block_size);
     const float pow4_normalization_factor = pow(norm_factor, 4);
 
+    // Manage intermediate buffers
+    _memory_group.manage(&_score);
+
     // Set/init Harris Score kernel accordingly with block_size
     _harris_score.configure(&_gx, &_gy, &_score, block_size, pow4_normalization_factor, threshold, sensitivity, border_mode == BorderMode::UNDEFINED);
 
@@ -123,26 +131,35 @@ void CLHarrisCorners::configure(ICLImage *input, float threshold, float min_dist
     _border_gx.configure(&_gx, _harris_score.border_size(), border_mode, PixelValue(constant_border_value));
     _border_gy.configure(&_gy, _harris_score.border_size(), border_mode, PixelValue(constant_border_value));
 
+    // Allocate intermediate buffers
+    _gx.allocator()->allocate();
+    _gy.allocator()->allocate();
+
+    // Manage intermediate buffers
+    _memory_group.manage(&_nonmax);
+
     // Init non-maxima suppression function
     _non_max_suppr.configure(&_score, &_nonmax, border_mode);
 
+    // Allocate intermediate buffers
+    _score.allocator()->allocate();
+
     // Init corner candidates kernel
     _candidates.configure(&_nonmax, _corners_list.get(), &_num_corner_candidates);
 
-    // Init euclidean distance
-    _sort_euclidean.configure(_corners_list.get(), _corners, &_num_corner_candidates, min_dist);
-
     // Allocate intermediate buffers
-    _gx.allocator()->allocate();
-    _gy.allocator()->allocate();
-    _score.allocator()->allocate();
     _nonmax.allocator()->allocate();
+
+    // Init euclidean distance
+    _sort_euclidean.configure(_corners_list.get(), _corners, &_num_corner_candidates, min_dist);
 }
 
 void CLHarrisCorners::run()
 {
     ARM_COMPUTE_ERROR_ON_MSG(_sobel == nullptr, "Unconfigured function");
 
+    _memory_group.acquire();
+
     // Init to 0 number of corner candidates
     _num_corner_candidates = 0;
 
@@ -167,4 +184,6 @@ void CLHarrisCorners::run()
     _corners->map(CLScheduler::get().queue(), true);
     Scheduler::get().schedule(&_sort_euclidean, Window::DimY);
     _corners->unmap(CLScheduler::get().queue());
+
+    _memory_group.release();
 }
diff --git a/src/runtime/CL/functions/CLSobel5x5.cpp b/src/runtime/CL/functions/CLSobel5x5.cpp
index 098b546c1a..d4bc85524e 100644
--- a/src/runtime/CL/functions/CLSobel5x5.cpp
+++ b/src/runtime/CL/functions/CLSobel5x5.cpp
@@ -33,8 +33,8 @@
 
 using namespace arm_compute;
 
-CLSobel5x5::CLSobel5x5()
-    : _sobel_hor(), _sobel_vert(), _border_handler(), _tmp_x(), _tmp_y()
+CLSobel5x5::CLSobel5x5(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _sobel_hor(), _sobel_vert(), _border_handler(), _tmp_x(), _tmp_y()
 {
 }
 
@@ -51,6 +51,8 @@ void CLSobel5x5::configure(ICLTensor *input, ICLTensor *output_x, ICLTensor *out
     {
         _tmp_x.allocator()->init(tensor_info);
         _tmp_y.allocator()->init(tensor_info);
+        _memory_group.manage(&_tmp_x);
+        _memory_group.manage(&_tmp_y);
         _sobel_hor.configure(input, &_tmp_x, &_tmp_y, border_mode == BorderMode::UNDEFINED);
         _sobel_vert.configure(&_tmp_x, &_tmp_y, output_x, output_y, border_mode == BorderMode::UNDEFINED);
         _tmp_x.allocator()->allocate();
@@ -59,6 +61,7 @@ void CLSobel5x5::configure(ICLTensor *input, ICLTensor *output_x, ICLTensor *out
     else if(run_sobel_x)
     {
         _tmp_x.allocator()->init(tensor_info);
+        _memory_group.manage(&_tmp_x);
         _sobel_hor.configure(input, &_tmp_x, nullptr, border_mode == BorderMode::UNDEFINED);
         _sobel_vert.configure(&_tmp_x, nullptr, output_x, nullptr, border_mode == BorderMode::UNDEFINED);
         _tmp_x.allocator()->allocate();
@@ -66,6 +69,7 @@ void CLSobel5x5::configure(ICLTensor *input, ICLTensor *output_x, ICLTensor *out
     else if(run_sobel_y)
     {
         _tmp_y.allocator()->init(tensor_info);
+        _memory_group.manage(&_tmp_y);
         _sobel_hor.configure(input, nullptr, &_tmp_y, border_mode == BorderMode::UNDEFINED);
         _sobel_vert.configure(nullptr, &_tmp_y, nullptr, output_y, border_mode == BorderMode::UNDEFINED);
         _tmp_y.allocator()->allocate();
@@ -76,6 +80,11 @@ void CLSobel5x5::configure(ICLTensor *input, ICLTensor *output_x, ICLTensor *out
 void CLSobel5x5::run()
 {
     CLScheduler::get().enqueue(_border_handler, false);
+
+    _memory_group.acquire();
+
     CLScheduler::get().enqueue(_sobel_hor, false);
     CLScheduler::get().enqueue(_sobel_vert);
+
+    _memory_group.release();
 }
diff --git a/src/runtime/CL/functions/CLSobel7x7.cpp b/src/runtime/CL/functions/CLSobel7x7.cpp
index db84fa99ae..60830905df 100644
--- a/src/runtime/CL/functions/CLSobel7x7.cpp
+++ b/src/runtime/CL/functions/CLSobel7x7.cpp
@@ -33,8 +33,8 @@
 
 using namespace arm_compute;
 
-CLSobel7x7::CLSobel7x7()
-    : _sobel_hor(), _sobel_vert(), _border_handler(), _tmp_x(), _tmp_y()
+CLSobel7x7::CLSobel7x7(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _sobel_hor(), _sobel_vert(), _border_handler(), _tmp_x(), _tmp_y()
 {
 }
 
@@ -51,6 +51,8 @@ void CLSobel7x7::configure(ICLTensor *input, ICLTensor *output_x, ICLTensor *out
     {
         _tmp_x.allocator()->init(tensor_info);
         _tmp_y.allocator()->init(tensor_info);
+        _memory_group.manage(&_tmp_x);
+        _memory_group.manage(&_tmp_y);
         _sobel_hor.configure(input, &_tmp_x, &_tmp_y, border_mode == BorderMode::UNDEFINED);
         _sobel_vert.configure(&_tmp_x, &_tmp_y, output_x, output_y, border_mode == BorderMode::UNDEFINED);
         _tmp_x.allocator()->allocate();
@@ -59,6 +61,7 @@ void CLSobel7x7::configure(ICLTensor *input, ICLTensor *output_x, ICLTensor *out
     else if(run_sobel_x)
     {
         _tmp_x.allocator()->init(tensor_info);
+        _memory_group.manage(&_tmp_x);
         _sobel_hor.configure(input, &_tmp_x, nullptr, border_mode == BorderMode::UNDEFINED);
         _sobel_vert.configure(&_tmp_x, nullptr, output_x, nullptr, border_mode == BorderMode::UNDEFINED);
         _tmp_x.allocator()->allocate();
@@ -66,6 +69,7 @@ void CLSobel7x7::configure(ICLTensor *input, ICLTensor *output_x, ICLTensor *out
     else if(run_sobel_y)
     {
         _tmp_y.allocator()->init(tensor_info);
+        _memory_group.manage(&_tmp_y);
         _sobel_hor.configure(input, nullptr, &_tmp_y, border_mode == BorderMode::UNDEFINED);
         _sobel_vert.configure(nullptr, &_tmp_y, nullptr, output_y, border_mode == BorderMode::UNDEFINED);
         _tmp_y.allocator()->allocate();
@@ -76,6 +80,11 @@ void CLSobel7x7::configure(ICLTensor *input, ICLTensor *output_x, ICLTensor *out
 void CLSobel7x7::run()
 {
     CLScheduler::get().enqueue(_border_handler, false);
+
+    _memory_group.acquire();
+
     CLScheduler::get().enqueue(_sobel_hor, false);
     CLScheduler::get().enqueue(_sobel_vert);
+
+    _memory_group.release();
 }
diff --git a/src/runtime/NEON/functions/NEGaussian5x5.cpp b/src/runtime/NEON/functions/NEGaussian5x5.cpp
index f085975b1e..b010ca0104 100644
--- a/src/runtime/NEON/functions/NEGaussian5x5.cpp
+++ b/src/runtime/NEON/functions/NEGaussian5x5.cpp
@@ -57,9 +57,10 @@ void NEGaussian5x5::configure(ITensor *input, ITensor *output, BorderMode border
 
 void NEGaussian5x5::run()
 {
+    NEScheduler::get().schedule(&_border_handler, Window::DimZ);
+
     _memory_group.acquire();
 
-    NEScheduler::get().schedule(&_border_handler, Window::DimZ);
     NEScheduler::get().schedule(&_kernel_hor, Window::DimY);
     NEScheduler::get().schedule(&_kernel_vert, Window::DimY);