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

Adds support to:
-NEHarrisCorners
-NEFastCorners

Change-Id: I21eeb9961092d20f69e82f447fe1faa48b0f4ad8
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/88113
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

4 files changed, 45 insertions, 12 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEFastCorners.h b/arm_compute/runtime/NEON/functions/NEFastCorners.h
index d7c31750c5..5ecf0c2b19 100644
--- a/arm_compute/runtime/NEON/functions/NEFastCorners.h
+++ b/arm_compute/runtime/NEON/functions/NEFastCorners.h
@@ -31,9 +31,12 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/Array.h"
 #include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
+#include "arm_compute/runtime/MemoryGroup.h"
 #include "arm_compute/runtime/Tensor.h"
 
 #include <cstdint>
+#include <memory>
 
 namespace arm_compute
 {
@@ -51,7 +54,7 @@ class NEFastCorners : public IFunction
 {
 public:
     /** Constructor */
-    NEFastCorners();
+    NEFastCorners(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Initialize the function's source, destination, conv and border_mode.
      *
      * @param[in, out] input                 Source image. Data type supported: U8. (Written to only for @p border_mode != UNDEFINED)
@@ -68,6 +71,7 @@ public:
     void run() override;
 
 private:
+    MemoryGroup                     _memory_group;
     NEFastCornersKernel             _fast_corners_kernel;
     NEFillBorderKernel              _border_handler;
     NENonMaximaSuppression3x3Kernel _nonmax_kernel;
diff --git a/arm_compute/runtime/NEON/functions/NEHarrisCorners.h b/arm_compute/runtime/NEON/functions/NEHarrisCorners.h
index a709871153..a9a6786db5 100644
--- a/arm_compute/runtime/NEON/functions/NEHarrisCorners.h
+++ b/arm_compute/runtime/NEON/functions/NEHarrisCorners.h
@@ -31,6 +31,8 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/Array.h"
 #include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
+#include "arm_compute/runtime/MemoryGroup.h"
 #include "arm_compute/runtime/NEON/functions/NENonMaximaSuppression3x3.h"
 #include "arm_compute/runtime/Tensor.h"
 
@@ -63,7 +65,7 @@ public:
      *
      * Initialize _sobel, _harris_score and _corner_list to nullptr.
      */
-    NEHarrisCorners();
+    NEHarrisCorners(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Initialize the function's source, destination, conv and border_mode.
      *
      * @param[in, out] input                 Source image. Data type supported: U8. (Written to only for @p border_mode != UNDEFINED)
@@ -85,6 +87,7 @@ public:
     void run() override;
 
 private:
+    MemoryGroup                           _memory_group;          /**< Function's memory group */
     std::unique_ptr<IFunction>            _sobel;                 /**< Sobel function */
     std::unique_ptr<INEHarrisScoreKernel> _harris_score;          /**< Harris score kernel */
     NENonMaximaSuppression3x3             _non_max_suppr;         /**< Non-maxima suppression function */
diff --git a/src/runtime/NEON/functions/NEFastCorners.cpp b/src/runtime/NEON/functions/NEFastCorners.cpp
index 265041fc42..4137b1d131 100644
--- a/src/runtime/NEON/functions/NEFastCorners.cpp
+++ b/src/runtime/NEON/functions/NEFastCorners.cpp
@@ -35,8 +35,9 @@
 
 using namespace arm_compute;
 
-NEFastCorners::NEFastCorners()
-    : _fast_corners_kernel(),
+NEFastCorners::NEFastCorners(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)),
+      _fast_corners_kernel(),
       _border_handler(),
       _nonmax_kernel(),
       _fill_kernel(),
@@ -59,6 +60,7 @@ void NEFastCorners::configure(IImage *input, float threshold, bool nonmax_suppre
 
     TensorInfo tensor_info(input->info()->tensor_shape(), Format::U8);
     _output.allocator()->init(tensor_info);
+    _memory_group.manage(&_output);
 
     // If border is UNDEFINED _fast_corners_kernel will operate in xwindow (3,
     // width - 3) and ywindow (3, height -3) so the output image will leave the
@@ -75,6 +77,7 @@ void NEFastCorners::configure(IImage *input, float threshold, bool nonmax_suppre
     else
     {
         _suppressed.allocator()->init(tensor_info);
+        _memory_group.manage(&_suppressed);
         _nonmax_kernel.configure(&_output, &_suppressed, BorderMode::UNDEFINED == border_mode);
         _fill_kernel.configure(&_suppressed, 1 /* we keep all texels >0 */, corners);
 
@@ -90,6 +93,8 @@ void NEFastCorners::run()
 {
     NEScheduler::get().schedule(&_border_handler, Window::DimZ);
 
+    _memory_group.acquire();
+
     NEScheduler::get().schedule(&_fast_corners_kernel, Window::DimY);
 
     if(_non_max)
@@ -98,4 +103,6 @@ void NEFastCorners::run()
     }
 
     NEScheduler::get().schedule(&_fill_kernel, Window::DimY);
+
+    _memory_group.release();
 }
diff --git a/src/runtime/NEON/functions/NEHarrisCorners.cpp b/src/runtime/NEON/functions/NEHarrisCorners.cpp
index 7ec681dce9..25e28d2213 100644
--- a/src/runtime/NEON/functions/NEHarrisCorners.cpp
+++ b/src/runtime/NEON/functions/NEHarrisCorners.cpp
@@ -41,8 +41,9 @@
 
 using namespace arm_compute;
 
-NEHarrisCorners::NEHarrisCorners() // NOLINT
-    : _sobel(),
+NEHarrisCorners::NEHarrisCorners(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
+    : _memory_group(std::move(memory_manager)),
+      _sobel(),
       _harris_score(),
       _non_max_suppr(),
       _candidates(),
@@ -81,6 +82,10 @@ void NEHarrisCorners::configure(IImage *input, float threshold, float min_dist,
     _gx.allocator()->init(tensor_info_gxgy);
     _gy.allocator()->init(tensor_info_gxgy);
 
+    // Manage intermediate buffers
+    _memory_group.manage(&_gx);
+    _memory_group.manage(&_gy);
+
     TensorInfo tensor_info_score(shape, Format::F32);
     _score.allocator()->init(tensor_info_score);
     _nonmax.allocator()->init(tensor_info_score);
@@ -118,6 +123,9 @@ void NEHarrisCorners::configure(IImage *input, float threshold, float min_dist,
     // Normalization factor
     const float norm_factor = 1.0f / (255.0f * pow(4.0f, gradient_size / 2) * block_size);
 
+    // Manage intermediate buffers
+    _memory_group.manage(&_score);
+
     if(use_fp16)
     {
         switch(block_size)
@@ -180,26 +188,35 @@ void NEHarrisCorners::configure(IImage *input, float threshold, float min_dist,
     _border_gx.configure(&_gx, _harris_score->border_size(), border_mode, constant_border_value);
     _border_gy.configure(&_gy, _harris_score->border_size(), border_mode, constant_border_value);
 
+    // Allocate once all the configure methods have been called
+    _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 once all the configure methods have been called
+    _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 once all the configure methods have been called
-    _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 NEHarrisCorners::run()
 {
     ARM_COMPUTE_ERROR_ON_MSG(_sobel == nullptr, "Unconfigured function");
 
+    _memory_group.acquire();
+
     // Init to 0 number of corner candidates
     _num_corner_candidates = 0;
 
@@ -221,4 +238,6 @@ void NEHarrisCorners::run()
 
     // Run sort & euclidean distance
     NEScheduler::get().schedule(&_sort_euclidean, Window::DimY);
+
+    _memory_group.release();
 }