Remove padding in cpuPool2d NCHW

Remove padding from all cpuPool2d NCHW kernels (FP16,FP32 & Quantized)

Resolves: COMPMID-4728, COMPMID-4823

Signed-off-by: Freddie Liardet <frederick.liardet@arm.com>
Change-Id: Ida619f67cd6606b33828f2d9dee925aeb794cc50
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6358
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

8 files changed, 359 insertions, 357 deletions

diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 37ba9f93bf..b2b09825c1 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -675,8 +675,8 @@ public:
      * @param[in] stride_x   Stride, in elements, across x.
      * @param[in] stride_y   Stride, in elements, across y.
      * @param[in] pad_left   Padding across x on the left, in elements.
-     * @param[in] pad_top    Padding across y on the top, in elements.
      * @param[in] pad_right  Padding across x on the right, in elements.
+     * @param[in] pad_top    Padding across y on the top, in elements.
      * @param[in] pad_bottom Padding across y on the bottom, in elements.
      * @param[in] round      Dimensions rounding.
      */
diff --git a/src/cpu/kernels/CpuPool2dKernel.cpp b/src/cpu/kernels/CpuPool2dKernel.cpp
index d7fb75ee60..90ddefa599 100644
--- a/src/cpu/kernels/CpuPool2dKernel.cpp
+++ b/src/cpu/kernels/CpuPool2dKernel.cpp
@@ -239,7 +239,6 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const
 
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst, ITensorInfo *indices, const PoolingLayerInfo &pool_info,
                                                         unsigned int &num_elems_processed_per_iteration,
-                                                        BorderSize   &border_size,
                                                         int pool_size_x, int pool_size_y)
 {
     // dst auto inizialitation if not yet initialized
@@ -251,29 +250,22 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITenso
                                                                                         pool_info)))
                            .set_data_type(DataType::U32) /* we store the offset to the element */);
     }
-    const auto          data_layout                  = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout;
-    unsigned int        num_elems_read_per_iteration = 0;
-    unsigned int        num_elems_horizontal_window  = 0;
-    int                 pool_stride_x                = 0;
-    int                 pool_stride_y                = 0;
-    const int           idx_width                    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const int           idx_height                   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const int           src_width                    = src->dimension(idx_width);
-    const int           src_height                   = src->dimension(idx_height);
-    const PadStrideInfo pad_stride_info              = pool_info.pad_stride_info;
+    const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout;
+    ARM_COMPUTE_ERROR_ON(src->data_layout() != DataLayout::NCHW);
+
+    int                 pool_stride_x   = 0;
+    int                 pool_stride_y   = 0;
+    const int           idx_width       = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int           idx_height      = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+    const PadStrideInfo pad_stride_info = pool_info.pad_stride_info;
+
     std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride();
-    const int          pool_pad_right  = pad_stride_info.pad_right();
-    const int          pool_pad_top    = pad_stride_info.pad_top();
-    const int          pool_pad_left   = pad_stride_info.pad_left();
-    const int          pool_pad_bottom = pad_stride_info.pad_bottom();
-    const bool         is_square       = pool_size_x == pool_size_y;
-    const unsigned int pooled_w        = dst->dimension(idx_width);
-    const unsigned int pooled_h        = dst->dimension(idx_height);
+    const bool         is_square = pool_size_x == pool_size_y;
+    const unsigned int pooled_w  = dst->dimension(idx_width);
+    const unsigned int pooled_h  = dst->dimension(idx_height);
 
     //If it's not squared and optimized will be executed the MxN
-    num_elems_read_per_iteration      = 1;
     num_elems_processed_per_iteration = 1;
-    num_elems_horizontal_window       = 1;
 
     if(is_square)
     {
@@ -284,14 +276,10 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITenso
                 switch(pool_size_x)
                 {
                     case 2:
-                        num_elems_read_per_iteration      = 16;
                         num_elems_processed_per_iteration = (pool_stride_x == 2) ? 8 : 15;
-                        num_elems_horizontal_window       = (pool_stride_x == 2) ? 8 : 16;
                         break;
                     case 3:
-                        num_elems_read_per_iteration      = 16;
                         num_elems_processed_per_iteration = (pool_stride_x == 2) ? 7 : 14;
-                        num_elems_horizontal_window       = (pool_stride_x == 2) ? 8 : 16;
                         break;
                     default:
                         break;
@@ -299,36 +287,11 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITenso
                 break;
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
             case DataType::F16:
-                switch(pool_size_x)
-                {
-                    case 2:
-                    case 3:
-                        num_elems_read_per_iteration      = 4;
-                        num_elems_processed_per_iteration = 1;
-                        num_elems_horizontal_window       = 1;
-                        break;
-                    default:
-                        break;
-                }
+                num_elems_processed_per_iteration = 1;
                 break;
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
             case DataType::F32:
-                switch(pool_size_x)
-                {
-                    case 2:
-                        num_elems_read_per_iteration = 2;
-                        break;
-                    case 3:
-                        num_elems_read_per_iteration = 4; // We use vload4 for pooling3
-                        break;
-                    case 7:
-                        num_elems_read_per_iteration = 8; // We use vload8 for pooling7
-                        break;
-                    default:
-                        break;
-                }
                 num_elems_processed_per_iteration = 1;
-                num_elems_horizontal_window       = 1;
                 break;
             default:
                 ARM_COMPUTE_ERROR("Element size not supported");
@@ -338,47 +301,18 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITenso
 
     bool   window_changed = false;
     Window win{};
-    if(data_layout == DataLayout::NCHW)
-    {
-        // Number of iterations in X dimension
-        const int num_iterations_x = (pooled_w + num_elems_processed_per_iteration - 1) / num_elems_processed_per_iteration;
-        // Upper limit for the number of right/bottom border elements that are accessed
-        const int upper_bound_w = ((num_iterations_x - 1) * num_elems_processed_per_iteration * pool_stride_x - pool_pad_left + num_elems_read_per_iteration) - src_width;
-        const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_top + pool_size_y) - src_height;
-        border_size             = BorderSize(pool_pad_top, pool_pad_right, pool_pad_bottom, pool_pad_left);
-        border_size.right       = std::max(upper_bound_w, pool_pad_right);
-        border_size.bottom      = std::max(upper_bound_h, pool_pad_bottom);
-        TensorShape dst_shape{ src->tensor_shape() };
-        dst_shape.set(0, pooled_w);
-        dst_shape.set(1, pooled_h);
-        TensorInfo dst_info(src->clone()->set_tensor_shape(dst_shape));
-        win = calculate_max_window(dst_info, Steps(num_elems_processed_per_iteration));
-        AccessWindowStatic     src_access(src, -pool_pad_left, -pool_pad_top, ceil_to_multiple(src_width + border_size.right, pool_size_x), src_height + border_size.bottom);
-        AccessWindowHorizontal dst_access(dst, 0, num_elems_horizontal_window);
-        if(indices)
-        {
-            AccessWindowHorizontal indices_access(indices, 0, num_elems_horizontal_window);
-            window_changed = update_window_and_padding(win, src_access, dst_access, indices_access);
-        }
-        else
-        {
-            window_changed = update_window_and_padding(win, src_access, dst_access);
-        }
-        dst_access.set_valid_region(win, ValidRegion(Coordinates(), dst->tensor_shape()));
-
-        border_size = src->padding();
-    }
+    // Upper limit for the number of right/bottom border elements that are accessed
+    TensorShape dst_shape{ src->tensor_shape() };
+    dst_shape.set(0, pooled_w);
+    dst_shape.set(1, pooled_h);
+    TensorInfo dst_info(src->clone()->set_tensor_shape(dst_shape));
+    win = calculate_max_window(dst_info, Steps(num_elems_processed_per_iteration));
 
     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
     return std::make_pair(err, win);
 }
 } // namespace
 
-BorderSize CpuPool2dKernel::border_size() const
-{
-    return _border_size;
-}
-
 void CpuPool2dKernel::configure(ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
@@ -419,7 +353,7 @@ void CpuPool2dKernel::configure(ITensorInfo *src, ITensorInfo *dst, const Poolin
     {
         // Configure kernel window
         auto win_config = validate_and_configure_window(src, dst, indices, pool_info, _num_elems_processed_per_iteration,
-                                                        _border_size, pool_size.x(), pool_size.y());
+                                                        pool_size.x(), pool_size.y());
         ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
         ICpuKernel::configure(win_config.second);
     }
@@ -430,7 +364,6 @@ Status CpuPool2dKernel::validate(const ITensorInfo *src, const ITensorInfo *dst,
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src);
 
     unsigned int num_elems_processed_per_iteration = 0;
-    BorderSize   border_size(0);
 
     const bool is_global_pooling = pool_info.is_global_pooling;
 
@@ -444,7 +377,7 @@ Status CpuPool2dKernel::validate(const ITensorInfo *src, const ITensorInfo *dst,
 
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, pool_info, indices, Size2D(pool_size_x, pool_size_y)));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(),
-                                                              (indices) ? indices->clone().get() : nullptr, pool_info, num_elems_processed_per_iteration, border_size,
+                                                              (indices) ? indices->clone().get() : nullptr, pool_info, num_elems_processed_per_iteration,
                                                               pool_size_x, pool_size_y)
                                 .first);
 
diff --git a/src/cpu/kernels/CpuPool2dKernel.h b/src/cpu/kernels/CpuPool2dKernel.h
index 70fe52d29c..aedeb7fbe9 100644
--- a/src/cpu/kernels/CpuPool2dKernel.h
+++ b/src/cpu/kernels/CpuPool2dKernel.h
@@ -60,7 +60,6 @@ public:
 
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
-    BorderSize  border_size() const override;
     const char *name() const override;
 
 private:
@@ -70,7 +69,6 @@ private:
     PoolingLayerInfo _pool_info{};
     DataLayout       _data_layout{ DataLayout::UNKNOWN };
     unsigned int     _num_elems_processed_per_iteration{ 0 };
-    BorderSize       _border_size{ 0 };
     Size2D           _pool_size{};
     int              _pool_stride_x{};
     PoolingKernelPtr _run_method{ nullptr };
diff --git a/src/cpu/kernels/pool2d/neon/nchw/all.cpp b/src/cpu/kernels/pool2d/neon/nchw/all.cpp
index 3ca7701087..109fc1b283 100644
--- a/src/cpu/kernels/pool2d/neon/nchw/all.cpp
+++ b/src/cpu/kernels/pool2d/neon/nchw/all.cpp
@@ -28,18 +28,55 @@
 #include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
 #include "src/core/helpers/WindowHelpers.h"
 #include "src/cpu/kernels/pool2d/neon/list.h"
+#include <limits>
 
 #ifdef ENABLE_NCHW_KERNELS
 namespace arm_compute
 {
 namespace cpu
 {
+#define READ_2_RIGHT_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval) \
+    (x == width + pad_left - 1) ? vset_lane_f32(*(ptr), vdup_n_f32(fval), 0) : vld1_f32(ptr)
+#define READ_2_LEFT_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval) \
+    (x == pad_left - 1) ? vset_lane_f32(*(1 + ptr), vdup_n_f32(fval), 1) : READ_2_RIGHT_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval)
+#define READ_2_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval) \
+    ((y < pad_top) || (x < pad_left - 1) || (y >= height + pad_top) || (x > width + pad_left - 1)) ? vdup_n_f32(fval) : READ_2_LEFT_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval)
+
+#define READ_4_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval)           \
+    vcombine_f32(READ_2_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval), \
+                 READ_2_BOUNDARY_AWARE(height, width, pad_left, pad_top, (x + 2), y, (ptr + 2), fval))
+
+float32x4x2_t read_8_boundary_aware(int height, int width, int pad_left, int pad_top, int x, int y, const float *ptr, float fval)
+{
+    float32x4x2_t vec;
+    vec.val[0] = READ_4_BOUNDARY_AWARE(height, width, pad_left, pad_top, x, y, ptr, fval);
+    vec.val[1] = READ_4_BOUNDARY_AWARE(height, width, pad_left, pad_top, (x + 4), y, (ptr + 4), fval);
+    return vec;
+}
+
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+float16x4_t read_4_boundary_aware_fp16(int srcw, int srch, int pad_l, int pad_t, int x, int y, const float16_t *ptr, float16_t fval)
+{
+    float16_t  vec[4];
+    const bool row_in_bounds((y >= pad_t) && (y < (srch + pad_t)));
+    for(int i = 0; i < 4; i++)
+    {
+        if(row_in_bounds && (x + i >= pad_l) && (x + i < (srcw + pad_l)))
+        {
+            vec[i] = *(ptr + i);
+        }
+        else
+        {
+            vec[i] = fval;
+        }
+    }
+    return wrapper::vload(vec);
+}
+
 void pooling3_fp16_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &pool_info, const Window &window_src, const Window &window)
 {
     ARM_COMPUTE_UNUSED(dst1);
-    ARM_COMPUTE_UNUSED(pool_info.pool_type);
-    ARM_COMPUTE_UNUSED(pool_info.exclude_padding);
 
     Iterator in(src, window_src);
     Iterator out(dst0, window);
@@ -52,19 +89,29 @@ void pooling3_fp16_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
     int                 pool_stride_x   = 0;
     int                 pool_stride_y   = 0;
     std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride();
-    const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
-    const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
-
+    const int                  src_w          = src->info()->dimension(0);
+    const int                  src_h          = src->info()->dimension(1);
+    const int                  upper_bound_w  = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right);
+    const int                  upper_bound_h  = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    constexpr float16_t        fp16_min       = -100.0f;
+    const float16_t            fill_value     = (pool_info.pool_type == PoolingType::MAX) ? fp16_min : 0.f;
     const unsigned char *const src_top_ptr    = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top)));
     const unsigned char *const src_middle_ptr = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1));
     const unsigned char *const src_bottom_ptr = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 2));
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        float16x4_t top_data    = vld1_f16(reinterpret_cast<const float16_t *>(src_top_ptr + in.offset()));
-        float16x4_t middle_data = vld1_f16(reinterpret_cast<const float16_t *>(src_middle_ptr + in.offset()));
-        float16x4_t bottom_data = vld1_f16(reinterpret_cast<const float16_t *>(src_bottom_ptr + in.offset()));
-        float16x4_t res         = {};
+        const auto  x_val    = id.x() * pool_stride_x;
+        const auto  y_val_0  = id.y() * pool_stride_y;
+        const auto  y_val_1  = (id.y() * pool_stride_y) + 1;
+        const auto  y_val_2  = (id.y() * pool_stride_y) + 2;
+        float16x4_t top_data = read_4_boundary_aware_fp16(src_w, src_h, pool_pad_left, pool_pad_top,
+                                                          x_val, y_val_0, reinterpret_cast<const float16_t *>(src_top_ptr + in.offset()), fill_value);
+        float16x4_t middle_data = read_4_boundary_aware_fp16(src_w, src_h, pool_pad_left, pool_pad_top,
+                                                             x_val, y_val_1, reinterpret_cast<const float16_t *>(src_middle_ptr + in.offset()), fill_value);
+        float16x4_t bottom_data = read_4_boundary_aware_fp16(src_w, src_h, pool_pad_left, pool_pad_top,
+                                                             x_val, y_val_2, reinterpret_cast<const float16_t *>(src_bottom_ptr + in.offset()), fill_value);
+        float16x4_t res = {};
 
         // Get power of 2 in case of l2 pooling
         if(pool_info.pool_type == PoolingType::L2)
@@ -88,7 +135,7 @@ void pooling3_fp16_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
         else
         {
             const float16x4_t max_data = vmax_f16(vmax_f16(top_data, bottom_data), middle_data);
-            res                        = vpmax_f16(vset_lane_f16(-std::numeric_limits<float>::max(), max_data, 3), max_data);
+            res                        = vpmax_f16(vset_lane_f16(fp16_min, max_data, 3), max_data);
             res                        = vpmax_f16(res, res);
         }
 
@@ -120,6 +167,25 @@ f16_to_f32(float32x2_t in)
 }
 
 template <typename T>
+auto read_2_boundary_aware(int srcw, int srch, int pad_l, int pad_t, int x, int y, const T *ptr, T fval)
+{
+    T          vec[2];
+    const bool row_in_bounds((y >= pad_t) && (y < (srch + pad_t)));
+    for(int i = 0; i < 2; i++)
+    {
+        if(row_in_bounds && (x + i >= pad_l) && (x + i < (srcw + pad_l)))
+        {
+            vec[i] = *(ptr + i);
+        }
+        else
+        {
+            vec[i] = fval;
+        }
+    }
+    return wrapper::vload(vec);
+}
+
+template <typename T>
 void pooling2_nchw_maxpool_indices(const ITensor *src, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &pool_info, const Window &window_src, const Window &window)
 {
     Iterator  in(src, window_src);
@@ -130,16 +196,25 @@ void pooling2_nchw_maxpool_indices(const ITensor *src, ITensor *dst0, ITensor *d
     int       pool_stride_x = 0;
     int       pool_stride_y = 0;
     std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride();
+    const int            src_w          = src->info()->dimension(0);
+    const int            src_h          = src->info()->dimension(1);
     const uint8_t *const src_top_ptr    = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top)));
     const uint8_t *const src_bottom_ptr = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1));
     const int            pad_left       = src->info()->padding().left;
     const int            pad_right      = src->info()->padding().right;
     const int            in_stride_y    = static_cast<int>(src->info()->strides_in_bytes().y());
+    constexpr T          float_min      = -100.0f;
+    const T              fill_value     = (pool_info.pool_type == PoolingType::MAX) ? float_min : 0.f;
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        auto        top_data        = wrapper::vload(reinterpret_cast<const T *>(src_top_ptr + in.offset()));
-        auto        bottom_data     = wrapper::vload(reinterpret_cast<const T *>(src_bottom_ptr + in.offset()));
+        const auto x_val    = id.x() * pool_stride_x;
+        const auto y_val_0  = id.y() * pool_stride_y;
+        const auto y_val_1  = (id.y() * pool_stride_y) + 1;
+        auto       top_data = read_2_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_top,
+                                                    x_val, y_val_0, reinterpret_cast<const T *>(src_top_ptr + in.offset()), fill_value);
+        auto bottom_data = read_2_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_top,
+                                                 x_val, y_val_1, reinterpret_cast<const T *>(src_bottom_ptr + in.offset()), fill_value);
         float32x2_t top_data_f32    = f16_to_f32<T>(top_data);
         float32x2_t bottom_data_f32 = f16_to_f32<T>(bottom_data);
 
@@ -180,17 +255,29 @@ void pooling2_fp16_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
         const int     pool_pad_bottom = pool_info.pad_stride_info.pad_bottom();
         int           pool_stride_x, pool_stride_y = 0;
         std::tie(pool_stride_x, pool_stride_y)     = pool_info.pad_stride_info.stride();
-        const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
-        const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+        const int           src_w         = src->info()->dimension(0);
+        const int           src_h         = src->info()->dimension(1);
+        const int           upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right);
+        const int           upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+        constexpr float16_t fp16_min      = -100.0f;
+        const float16_t     fill_value    = (pool_info.pool_type == PoolingType::MAX) ? fp16_min : 0.0f;
 
         const unsigned char *const src_top_ptr    = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top)));
         const unsigned char *const src_bottom_ptr = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1));
 
         execute_window_loop(window, [&](const Coordinates & id)
         {
-            float16x4_t top_data    = vld1_f16(reinterpret_cast<const float16_t *>(src_top_ptr + in.offset()));
-            float16x4_t bottom_data = vld1_f16(reinterpret_cast<const float16_t *>(src_bottom_ptr + in.offset()));
-            float16x4_t res         = {};
+            const auto in_top_ptr    = reinterpret_cast<const float16_t *>(src_top_ptr + in.offset());
+            const auto in_bottom_ptr = reinterpret_cast<const float16_t *>(src_bottom_ptr + in.offset());
+
+            const auto  x_val    = id.x() * pool_stride_x;
+            const auto  y_val_0  = id.y() * pool_stride_y;
+            const auto  y_val_1  = (id.y() * pool_stride_y) + 1;
+            float16x4_t top_data = read_4_boundary_aware_fp16(src_w, src_h, pool_pad_left, pool_pad_top,
+                                                              x_val, y_val_0, in_top_ptr, fill_value);
+            float16x4_t bottom_data = read_4_boundary_aware_fp16(src_w, src_h, pool_pad_left, pool_pad_top,
+                                                                 x_val, y_val_1, in_bottom_ptr, fill_value);
+            float16x4_t res = {};
 
             // Get power of 2 in case of l2 pooling
             if(pool_info.pool_type == PoolingType::L2)
@@ -242,48 +329,35 @@ void poolingMxN_fp16_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1,
     int       pool_stride_x   = 0;
     int       pool_stride_y   = 0;
     std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride();
-    const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
-    const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    const int           src_w         = src->info()->dimension(0);
+    const int           src_h         = src->info()->dimension(1);
+    const int           upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right);
+    const int           upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    constexpr float16_t fp16_min      = -100.0f;
+    const float16_t     fill_value    = (pool_info.pool_type == PoolingType::MAX) ? fp16_min : 0.0f;
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        float16_t   res  = 0.0f;
-        float16x8_t vres = vdupq_n_f16(0.0f);
+        float16_t res = 0.0f;
 
         if(pool_info.pool_type != PoolingType::MAX)
         {
             // Calculate scale
-            const float scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x,
-                                                    pool_stride_y);
+            const float16_t scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x,
+                                                        pool_stride_y);
 
             // Perform pooling
-
             for(int y = 0; y < pool_size_y; ++y)
             {
-                int x = 0;
-                for(; x <= (pool_size_x - 8); x += 8)
+                for(int x = 0; x < pool_size_x; ++x)
                 {
-                    const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                                                           (src->info()->strides_in_bytes().y())));
-
-                    // Get power of 2 in case of l2 pooling and accumulate
-                    if(pool_info.pool_type == PoolingType::L2)
-                    {
-                        vres = vaddq_f16(vres, vmulq_f16(data, data));
-                    }
-                    else
-                    {
-                        vres = vaddq_f16(vres, data);
-                    }
-                }
+                    const auto ptr = reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x())
+                                                                         + (y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().y()));
 
-                // Leftover for loop
-                for(; x < pool_size_x; ++x)
-                {
-                    float16_t data = *(reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x())
-                                                                           + (y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().y())));
+                    const int idx  = x + id.x() * pool_stride_x - pool_pad_left;
+                    const int idy  = y + id.y() * pool_stride_y - pool_pad_top;
+                    float16_t data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *ptr;
 
-                    // Get power of 2 in case of l2 pooling
                     if(pool_info.pool_type == PoolingType::L2)
                     {
                         data *= data;
@@ -293,45 +367,26 @@ void poolingMxN_fp16_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1,
                 }
             }
 
-            // Reduction
-            float16x4_t tmp = vpadd_f16(vget_high_f16(vres), vget_low_f16(vres));
-            res += vget_lane_f16(tmp, 0);
-            res += vget_lane_f16(tmp, 1);
-            res += vget_lane_f16(tmp, 2);
-            res += vget_lane_f16(tmp, 3);
-
             // Divide by scale
             res *= scale;
         }
-        else
+        else // if max pooling
         {
-            float16x8_t vres = vdupq_n_f16(std::numeric_limits<float>::lowest());
-            res              = std::numeric_limits<float>::lowest();
+            res = fp16_min;
 
             for(int y = 0; y < pool_size_y; ++y)
             {
-                int x = 0;
-                for(; x <= (pool_size_x - 8); x += 8)
+                for(int x = 0; x < pool_size_x; ++x)
                 {
-                    const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                                                           (src->info()->strides_in_bytes().y())));
-                    vres                   = vmaxq_f16(vres, data);
-                }
+                    const auto ptr = reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x())
+                                                                         + (y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().y()));
 
-                // Leftover for loop
-                for(; x < pool_size_x; ++x)
-                {
-                    const float16_t data = *(reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x())
-                                                                                 + (y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().y())));
-                    res = std::max(res, data);
+                    const int idx  = x + id.x() * pool_stride_x - pool_pad_left;
+                    const int idy  = y + id.y() * pool_stride_y - pool_pad_top;
+                    float16_t data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *ptr;
+                    res            = std::max(res, data);
                 }
             }
-
-            float16x4_t tmp = vpmax_f16(vget_high_f16(vres), vget_low_f16(vres));
-            res             = std::max(res, vget_lane_f16(tmp, 0));
-            res             = std::max(res, vget_lane_f16(tmp, 1));
-            res             = std::max(res, vget_lane_f16(tmp, 2));
-            res             = std::max(res, vget_lane_f16(tmp, 3));
         }
 
         // Calculate square-root in case of l2 pooling
@@ -362,8 +417,11 @@ void poolingMxN_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1,
     int       pool_stride_x   = 0;
     int       pool_stride_y   = 0;
     std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride();
-    const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
-    const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    const int   src_w         = src->info()->dimension(0);
+    const int   src_h         = src->info()->dimension(1);
+    const int   upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right);
+    const int   upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    const float fill_value    = (pool_info.pool_type == PoolingType::MAX) ? -std::numeric_limits<float>::max() : 0.0f;
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
@@ -372,38 +430,21 @@ void poolingMxN_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1,
         if(pool_info.pool_type != PoolingType::MAX)
         {
             // Calculate scale
-            const float scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x,
-                                                    pool_stride_y);
+            const float scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h,
+                                                    pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y);
 
             // Perform pooling
-            float32x4_t vres = vdupq_n_f32(0.0f);
-
             for(int y = 0; y < pool_size_y; ++y)
             {
-                int x = 0;
-                for(; x <= (pool_size_x - 4); x += 4)
+                for(int x = 0; x < pool_size_x; ++x)
                 {
-                    const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                                                       (src->info()->strides_in_bytes().y())));
+                    const auto ptr = reinterpret_cast<const float *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x())
+                                                                     + (y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().y()));
 
-                    // Get power of 2 in case of l2 pooling and accumulate
-                    if(pool_info.pool_type == PoolingType::L2)
-                    {
-                        vres = vmlaq_f32(vres, data, data);
-                    }
-                    else
-                    {
-                        vres = vaddq_f32(vres, data);
-                    }
-                }
+                    const int idx  = x + id.x() * pool_stride_x - pool_pad_left;
+                    const int idy  = y + id.y() * pool_stride_y - pool_pad_top;
+                    float     data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *ptr;
 
-                // Leftover for loop
-                for(; x < pool_size_x; ++x)
-                {
-                    float data = *(reinterpret_cast<const float *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                                   (src->info()->strides_in_bytes().y())));
-
-                    // Get power of 2 in case of l2 pooling
                     if(pool_info.pool_type == PoolingType::L2)
                     {
                         data *= data;
@@ -413,51 +454,26 @@ void poolingMxN_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1,
                 }
             }
 
-#if defined(__aarch64__)
-            // Reduction operation available on 64 bit architectures only
-            res += vaddvq_f32(vres);
-#else  // __aarch64__
-            // Reduction
-            float32x2_t tmp = vpadd_f32(vget_high_f32(vres), vget_low_f32(vres));
-            tmp             = vpadd_f32(tmp, tmp);
-
-            res += vget_lane_f32(tmp, 0);
-#endif // __aarch64__
             // Divide by scale
             res *= scale;
         }
-        else
+        else // if max pooling
         {
-            float32x4_t vres = vdupq_n_f32(std::numeric_limits<float>::lowest());
-            res              = std::numeric_limits<float>::lowest();
+            res = std::numeric_limits<float>::lowest();
 
             for(int y = 0; y < pool_size_y; ++y)
             {
-                int x = 0;
-                for(; x <= (pool_size_x - 4); x += 4)
+                for(int x = 0; x < pool_size_x; ++x)
                 {
-                    const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                                                       (src->info()->strides_in_bytes().y())));
-                    vres                   = vmaxq_f32(vres, data);
-                }
+                    const auto ptr = reinterpret_cast<const float *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x())
+                                                                     + (y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().y()));
 
-                // Leftover for loop
-                for(; x < pool_size_x; ++x)
-                {
-                    const float data = *(reinterpret_cast<const float *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                                         (src->info()->strides_in_bytes().y())));
-                    res              = std::max(res, data);
+                    const int idx  = x + id.x() * pool_stride_x - pool_pad_left;
+                    const int idy  = y + id.y() * pool_stride_y - pool_pad_top;
+                    float     data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *ptr;
+                    res            = std::max(res, data);
                 }
             }
-#if defined(__aarch64__)
-            // Reduction operation available on 64 bit architectures only
-            res = std::max(vmaxvq_f32(vres), res);
-#else  // __aarch64__
-            float32x2_t tmp = vpmax_f32(vget_high_f32(vres), vget_low_f32(vres));
-            tmp             = vpmax_f32(tmp, tmp);
-
-            res = std::max(res, vget_lane_f32(tmp, 0));
-#endif // __aarch64__
         }
 
         // Calculate square-root in case of l2 pooling
@@ -490,20 +506,28 @@ void pooling2_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
         int           pool_stride_x   = 0;
         int           pool_stride_y   = 0;
         std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride();
-        const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
-        const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+        const int   src_w         = src->info()->dimension(0);
+        const int   src_h         = src->info()->dimension(1);
+        const int   upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right);
+        const int   upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+        const float fill_value    = (pool_info.pool_type == PoolingType::MAX) ? -std::numeric_limits<float>::max() : 0.0f;
 
         const uint8_t *const src_top_ptr    = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top)));
         const uint8_t *const src_bottom_ptr = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1));
 
         execute_window_loop(window, [&](const Coordinates & id)
         {
-            const auto  in_top_ptr    = reinterpret_cast<const float *>(src_top_ptr + in.offset());
-            const auto  in_bottom_ptr = reinterpret_cast<const float *>(src_bottom_ptr + in.offset());
-            float32x2_t top_data      = vld1_f32(in_top_ptr);
-            float32x2_t bottom_data   = vld1_f32(in_bottom_ptr);
-            float32x2_t res           = {};
-            float       final_res     = 0;
+            const auto in_top_ptr    = reinterpret_cast<const float *>(src_top_ptr + in.offset());
+            const auto in_bottom_ptr = reinterpret_cast<const float *>(src_bottom_ptr + in.offset());
+
+            const auto  x_val       = id.x() * pool_stride_x;
+            const auto  y_val_0     = id.y() * pool_stride_y;
+            const auto  y_val_1     = (id.y() * pool_stride_y) + 1;
+            auto        top_data    = READ_2_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_0, in_top_ptr, fill_value);
+            auto        bottom_data = READ_2_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_1, in_bottom_ptr, fill_value);
+            float32x2_t res         = {};
+            float       final_res   = 0;
+
             // Get power of 2 in case of l2 pooling
             if(pool_info.pool_type == PoolingType::L2)
             {
@@ -556,8 +580,11 @@ void pooling3_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
     int                 pool_stride_x   = 0;
     int                 pool_stride_y   = 0;
     std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride();
-    const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
-    const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    const int   src_w         = src->info()->dimension(0);
+    const int   src_h         = src->info()->dimension(1);
+    const int   upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right);
+    const int   upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    const float fill_value    = (pool_info.pool_type == PoolingType::MAX) ? -std::numeric_limits<float>::max() : 0.0f;
 
     const uint8_t *const src_top_ptr    = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top)));
     const uint8_t *const src_middle_ptr = src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1));
@@ -565,11 +592,20 @@ void pooling3_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        float32x4_t top_data    = vld1q_f32(reinterpret_cast<const float *>(src_top_ptr + in.offset()));
-        float32x4_t middle_data = vld1q_f32(reinterpret_cast<const float *>(src_middle_ptr + in.offset()));
-        float32x4_t bottom_data = vld1q_f32(reinterpret_cast<const float *>(src_bottom_ptr + in.offset()));
-        float32x2_t res         = {};
-        float       final_res   = 0;
+        const auto in_top_ptr    = reinterpret_cast<const float *>(src_top_ptr + in.offset());
+        const auto in_middle_ptr = reinterpret_cast<const float *>(src_middle_ptr + in.offset());
+        const auto in_bottom_ptr = reinterpret_cast<const float *>(src_bottom_ptr + in.offset());
+
+        const auto x_val       = id.x() * pool_stride_x;
+        const auto y_val_0     = id.y() * pool_stride_y;
+        const auto y_val_1     = (id.y() * pool_stride_y) + 1;
+        const auto y_val_2     = (id.y() * pool_stride_y) + 2;
+        auto       top_data    = READ_4_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_0, in_top_ptr, fill_value);
+        auto       middle_data = READ_4_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_1, in_middle_ptr, fill_value);
+        auto       bottom_data = READ_4_BOUNDARY_AWARE(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val_2, in_bottom_ptr, fill_value);
+
+        float32x2_t res       = {};
+        float       final_res = 0;
 
         // Get power of 2 in case of l2 pooling
         if(pool_info.pool_type == PoolingType::L2)
@@ -625,8 +661,11 @@ void pooling7_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
     int                 pool_stride_x   = 0;
     int                 pool_stride_y   = 0;
     std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride();
-    const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
-    const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    const int   src_w         = src->info()->dimension(0);
+    const int   src_h         = src->info()->dimension(1);
+    const int   upper_bound_w = src_w + (pool_info.exclude_padding ? 0 : pool_pad_right);
+    const int   upper_bound_h = src_h + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    const float fill_value    = (pool_info.pool_type == PoolingType::MAX) ? -std::numeric_limits<float>::max() : 0.0f;
 
     std::array<const uint8_t *, pool_size> src_ptrs{ {} };
     for(int i = 0; i < pool_size; ++i)
@@ -636,8 +675,15 @@ void pooling7_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
+        auto in_ptr = reinterpret_cast<const float *>(src_ptrs[0] + in.offset());
+
+        auto          x_val = id.x() * pool_stride_x;
+        auto          y_val = id.y() * pool_stride_y;
+        float32x4x2_t data  = read_8_boundary_aware(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val, in_ptr, fill_value);
+
         float32x2_t res       = {};
         float       final_res = 0.f;
+
         if(pool_info.pool_type != PoolingType::MAX)
         {
             // Calculate scale
@@ -645,8 +691,6 @@ void pooling7_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
                                               pool_stride_y);
             const float32x2_t scale_v = vdup_n_f32(scale);
 
-            // Perform pooling
-            float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(src_ptrs[0] + in.offset()));
             // Get power of 2 in case of l2 pooling
             if(pool_info.pool_type == PoolingType::L2)
             {
@@ -656,7 +700,11 @@ void pooling7_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
             float32x4_t sum_data = vaddq_f32(data.val[0], vsetq_lane_f32(0.f, data.val[1], 3));
             for(int i = 1; i < pool_size; ++i)
             {
-                data = vld2q_f32(reinterpret_cast<const float *>(src_ptrs[i] + in.offset()));
+                in_ptr = reinterpret_cast<const float *>(src_ptrs[i] + in.offset());
+
+                x_val = id.x() * pool_stride_x;
+                y_val = (id.y() * pool_stride_y) + i;
+                data  = read_8_boundary_aware(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val, in_ptr, fill_value);
                 // Get power of 2 in case of l2 pooling
                 if(pool_info.pool_type == PoolingType::L2)
                 {
@@ -671,14 +719,17 @@ void pooling7_fp32_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, P
         }
         else
         {
-            float32x4x2_t max_data = vld2q_f32(reinterpret_cast<const float *>(src_ptrs[0] + in.offset()));
             for(int i = 1; i < pool_size; ++i)
             {
-                const float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(src_ptrs[i] + in.offset()));
-                max_data                 = vmax2q_f32(max_data, data);
+                in_ptr = reinterpret_cast<const float *>(src_ptrs[i] + in.offset());
+
+                x_val              = id.x() * pool_stride_x;
+                y_val              = (id.y() * pool_stride_y) + i;
+                float32x4x2_t temp = read_8_boundary_aware(src_h, src_w, pool_pad_left, pool_pad_top, x_val, y_val, in_ptr, fill_value);
+                data               = vmax2q_f32(data, temp);
             }
-            res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data.val[1], 3)), vget_low_f32(max_data.val[1]));
-            res = vpmax_f32(res, vpmax_f32(vget_high_f32(max_data.val[0]), vget_low_f32(max_data.val[0])));
+            res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), data.val[1], 3)), vget_low_f32(data.val[1]));
+            res = vpmax_f32(res, vpmax_f32(vget_high_f32(data.val[0]), vget_low_f32(data.val[0])));
             res = vpmax_f32(res, res);
         }
         final_res = vget_lane_f32(res, 0);
diff --git a/src/cpu/kernels/pool2d/neon/quantized.h b/src/cpu/kernels/pool2d/neon/quantized.h
index a16960a205..386e043984 100644
--- a/src/cpu/kernels/pool2d/neon/quantized.h
+++ b/src/cpu/kernels/pool2d/neon/quantized.h
@@ -467,6 +467,63 @@ inline void scale_vector_q16x8(bool exclude_padding, TVec &v, const Coordinates
 }
 
 template <typename T>
+auto load16_boundary_aware(int srcw, int srch, int pad_l, int pad_r, int pad_t, int pad_b, int x, int y, const T *ptr, T fval)
+{
+    ARM_COMPUTE_UNUSED(pad_b, pad_r);
+    T vec[16];
+    //handle reading a row out of the tensor
+    const bool row_in_bounds((y >= pad_t) && (y < (srch + pad_t)));
+    for(int i = 0; i < 16; i++)
+    {
+        if(row_in_bounds && (x + i >= pad_l) && (x + i < (srcw + pad_l)))
+        {
+            vec[i] = *(ptr + i);
+        }
+        else
+        {
+            vec[i] = fval;
+        }
+    }
+    return wrapper::vloadq(vec);
+}
+
+template <typename T, typename V, bool deinterleave>
+inline void write16_boundary_aware(int x, int dst_w, const V &lower, const V &upper, T *ptr)
+{
+    if(deinterleave)
+    {
+        for(int i = 0; i < 8 && (i * 2 + x) < dst_w; ++i)
+        {
+            *(ptr + i * 2) = lower[i];
+        }
+        for(int i = 0; i < 8 && (i * 2 + x + 1) < dst_w; ++i)
+        {
+            *(ptr + 1 + i * 2) = upper[i];
+        }
+    }
+    else
+    {
+        for(int i = 0; i < 8 && (i + x) < dst_w; ++i)
+        {
+            *(ptr + i) = lower[i];
+        }
+        for(int i = 0; i < 8 && (i + x + 8) < dst_w; ++i)
+        {
+            *(ptr + i + 8) = upper[i];
+        }
+    }
+}
+
+template <typename T, typename V>
+inline void write8_boundary_aware(int x, int dst_w, const V &v, T *ptr)
+{
+    for(int i = 0; i < 8 && (i + x) < dst_w; ++i)
+    {
+        *(ptr + i) = v[i];
+    }
+}
+
+template <typename T>
 void pooling2_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &pool_info, const Window &window_src, const Window &window)
 {
     ARM_COMPUTE_UNUSED(dst1);
@@ -474,9 +531,8 @@ void pooling2_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *ds
     Iterator out(dst0, window);
 
     /** SIMD vector types */
-    using q8x8_t    = typename wrapper::traits::neon_vector<T, 8>::type;
-    using q8x16_t   = typename wrapper::traits::neon_vector<T, 16>::type;
-    using q8x8x2_t  = typename std::conditional<std::is_same<T, uint8_t>::value, uint8x8x2_t, int8x8x2_t>::type;
+    using q8x8_t  = typename wrapper::traits::neon_vector<T, 8>::type;
+    using q8x16_t = typename wrapper::traits::neon_vector<T, 16>::type;
     using q16_t     = typename wrapper::traits::promote_t<T>;
     using q16x4_t   = typename wrapper::traits::neon_vector<q16_t, 4>::type;
     using q16x8_t   = typename wrapper::traits::neon_vector<q16_t, 8>::type;
@@ -490,14 +546,11 @@ void pooling2_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *ds
     const int     pool_pad_left   = pool_info.pad_stride_info.pad_left();
     const int     pool_pad_bottom = pool_info.pad_stride_info.pad_bottom();
     std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride();
-    const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
-    const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
-
-    const T *const src_top_ptr    = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))));
-    const T *const src_bottom_ptr = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)));
-
-    const int scale_step_x = (pool_stride_x == 1) ? 2 : 1;
-
+    const int                     upper_bound_w        = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
+    const int                     upper_bound_h        = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
+    const T *const                src_top_ptr          = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))));
+    const T *const                src_bottom_ptr       = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)));
+    const int                     scale_step_x         = (pool_stride_x == 1) ? 2 : 1;
     const UniformQuantizationInfo src_qinfo            = src->info()->quantization_info().uniform();
     const UniformQuantizationInfo dst_qinfo            = dst0->info()->quantization_info().uniform();
     const bool                    have_different_qinfo = src_qinfo != dst_qinfo;
@@ -505,13 +558,25 @@ void pooling2_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *ds
     const float                   requant_scale  = dst_qinfo.scale / src_qinfo.scale;
     const int32_t                 requant_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / requant_scale);
     const UniformQuantizationInfo requant_qinfo  = UniformQuantizationInfo(requant_scale, requant_offset);
+    const int                     src_w          = src->info()->dimension(0);
+    const int                     src_h          = src->info()->dimension(1);
+    const int                     dst_w          = dst0->info()->dimension(0);
+
+    const T fill_value = (pool_info.pool_type == PoolingType::MAX) ? std::numeric_limits<T>::min() : T(0);
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        const auto top_data    = wrapper::vloadq(src_top_ptr + in.offset());
-        const auto bottom_data = wrapper::vloadq(src_bottom_ptr + in.offset());
-        q8x8_t     lower_res   = {};
-        q8x8_t     upper_res   = {};
+        const auto x_val   = id.x() * pool_stride_x;
+        const auto y_val_0 = id.y() * pool_stride_y;
+        const auto y_val_1 = (id.y() * pool_stride_y) + 1;
+
+        auto top_data = load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom,
+                                              x_val, y_val_0, reinterpret_cast<const T *>(src_top_ptr + in.offset()), fill_value);
+        auto bottom_data = load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom,
+                                                 x_val, y_val_1, reinterpret_cast<const T *>(src_bottom_ptr + in.offset()), fill_value);
+
+        q8x8_t lower_res = {};
+        q8x8_t upper_res = {};
 
         if(pool_info.pool_type != PoolingType::MAX)
         {
@@ -580,16 +645,15 @@ void pooling2_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *ds
             lower_res                  = wrapper::vgetlow(requantized_dst);
             upper_res                  = wrapper::vgethigh(requantized_dst);
         }
-
+        auto out_ptr = reinterpret_cast<T *>(out.ptr());
         // Store result
         if(pool_stride_x == 1)
         {
-            const q8x8x2_t res = { { lower_res, upper_res } };
-            wrapper::vstore(reinterpret_cast<T *>(out.ptr()), res);
+            write16_boundary_aware<T, q8x8_t, true>(id.x(), dst_w, lower_res, upper_res, out_ptr);
         }
         else
         {
-            wrapper::vstore(reinterpret_cast<T *>(out.ptr()), lower_res);
+            write8_boundary_aware<T, q8x8_t>(id.x(), dst_w, lower_res, out_ptr);
         }
     },
     in, out);
@@ -632,13 +696,27 @@ void pooling3_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *ds
     const T *const src_middle_ptr = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)));
     const T *const src_bottom_ptr = reinterpret_cast<const T *>(src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 2)));
 
+    const int src_w      = src->info()->dimension(0);
+    const int src_h      = src->info()->dimension(1);
+    const T   fill_value = (pool_info.pool_type == PoolingType::AVG) ? T(0) : std::numeric_limits<T>::min();
+    const int dst_w      = dst0->info()->dimension(0);
+
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        const auto top_data    = wrapper::vloadq(src_top_ptr + in.offset());
-        const auto middle_data = wrapper::vloadq(src_middle_ptr + in.offset());
-        const auto bottom_data = wrapper::vloadq(src_bottom_ptr + in.offset());
-        q8x8_t     fres        = {};
-        q8x16_t    fqres       = {};
+        const auto x_val   = id.x() * pool_stride_x;
+        const auto y_val_0 = id.y() * pool_stride_y;
+        const auto y_val_1 = (id.y() * pool_stride_y) + 1;
+        const auto y_val_2 = (id.y() * pool_stride_y) + 2;
+
+        auto top_data = load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom,
+                                              x_val, y_val_0, reinterpret_cast<const T *>(src_top_ptr + in.offset()), fill_value);
+        auto middle_data = load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom,
+                                                 x_val, y_val_1, reinterpret_cast<const T *>(src_middle_ptr + in.offset()), fill_value);
+        auto bottom_data = load16_boundary_aware(src_w, src_h, pool_pad_left, pool_pad_right, pool_pad_top, pool_pad_bottom,
+                                                 x_val, y_val_2, reinterpret_cast<const T *>(src_bottom_ptr + in.offset()), fill_value);
+
+        q8x8_t  fres  = {};
+        q8x16_t fqres = {};
 
         if(pool_info.pool_type == PoolingType::AVG)
         {
@@ -735,7 +813,7 @@ void pooling3_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *ds
             {
                 fqres = vrequantize_pooling<q8x8_t, q8x16_t>(wrapper::vgetlow(fqres), wrapper::vgethigh(fqres), requant_qinfo);
             }
-            wrapper::vstore(reinterpret_cast<T *>(out.ptr()), fqres);
+            write16_boundary_aware<T, q8x8_t, false>(id.x(), dst_w, wrapper::vgetlow(fqres), wrapper::vgethigh(fqres), reinterpret_cast<T *>(out.ptr()));
         }
         else
         {
@@ -743,7 +821,7 @@ void pooling3_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *ds
             {
                 fres = vrequantize_pooling<q8x8_t>(fres, requant_qinfo);
             }
-            wrapper::vstore(reinterpret_cast<T *>(out.ptr()), fres);
+            write8_boundary_aware<T, q8x8_t>(id.x(), dst_w, fres, reinterpret_cast<T *>(out.ptr()));
         }
     },
     in, out);
@@ -757,11 +835,8 @@ void poolingMxN_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *
     Iterator out(dst0, window);
 
     /** SIMD vector types */
-    using q8x8_t  = typename wrapper::traits::neon_vector<T, 8>::type;
-    using q16_t   = typename wrapper::traits::promote_t<T>;
-    using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type;
-    using q32_t   = typename wrapper::traits::promote_t<q16_t>;
-    using q32x4_t = typename wrapper::traits::neon_vector<q32_t, 4>::type;
+    using q16_t = typename wrapper::traits::promote_t<T>;
+    using q32_t = typename wrapper::traits::promote_t<q16_t>;
 
     const int pool_size_x     = pool_info.is_global_pooling ? src->info()->tensor_shape().x() : pool_info.pool_size.width;
     const int pool_size_y     = pool_info.is_global_pooling ? src->info()->tensor_shape().y() : pool_info.pool_size.height;
@@ -775,8 +850,13 @@ void poolingMxN_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *
     const int upper_bound_w = src->info()->dimension(0) + (pool_info.exclude_padding ? 0 : pool_pad_right);
     const int upper_bound_h = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
 
-    const UniformQuantizationInfo &src_qinfo = src->info()->quantization_info().uniform();
-    const UniformQuantizationInfo &dst_qinfo = dst0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo &src_qinfo        = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo &dst_qinfo        = dst0->info()->quantization_info().uniform();
+    const int                      src_w            = src->info()->dimension(0);
+    const int                      src_h            = src->info()->dimension(1);
+    const T                        fill_value       = (pool_info.pool_type == PoolingType::AVG) ? T(0) : std::numeric_limits<T>::min();
+    const int                      stridex_in_bytes = static_cast<int>(src->info()->strides_in_bytes().x());
+    const int                      stridey_in_bytes = static_cast<int>(src->info()->strides_in_bytes().y());
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
@@ -784,8 +864,7 @@ void poolingMxN_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *
 
         if(pool_info.pool_type != PoolingType::MAX)
         {
-            q32x4_t vres = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{});
-            q32_t   sres = 0;
+            q32_t sres = 0;
 
             // Calculate scale
             const float scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x,
@@ -794,61 +873,33 @@ void poolingMxN_quantized_neon_nchw(const ITensor *src, ITensor *dst0, ITensor *
             // Perform pooling
             for(int y = 0; y < pool_size_y; ++y)
             {
-                int x = 0;
-                for(; x <= (pool_size_x - 8); x += 8)
+                for(int x = 0; x < pool_size_x; ++x)
                 {
-                    const q8x8_t data = wrapper::vload(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                                                   (src->info()->strides_in_bytes().y())));
-
-                    const q16x8_t data_q16 = wrapper::vmovl(data);
-                    vres                   = wrapper::vadd(vres, wrapper::vaddl(wrapper::vgethigh(data_q16), wrapper::vgetlow(data_q16)));
-                }
+                    const auto in_ptr = reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * stridex_in_bytes + (y - pool_pad_top) * stridey_in_bytes);
 
-                // Leftover for loop
-                for(; x < pool_size_x; ++x)
-                {
-                    T data = *(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                           (src->info()->strides_in_bytes().y())));
+                    const int idx  = x + id.x() * pool_stride_x - pool_pad_left;
+                    const int idy  = y + id.y() * pool_stride_y - pool_pad_top;
+                    const T   data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *in_ptr;
                     sres += data;
                 }
             }
-
-            // Reduction
-            const auto tmp = wrapper::vpadd(wrapper::vgethigh(vres), wrapper::vgetlow(vres));
-            sres += wrapper::vgetlane(tmp, 0) + wrapper::vgetlane(tmp, 1);
-
             // Divide by scale
             res = static_cast<T>(support::cpp11::round(sres * scale));
         }
         else
         {
-            q8x8_t vres = wrapper::vdup_n(std::numeric_limits<T>::min(), wrapper::traits::vector_64_tag{});
-
             for(int y = 0; y < pool_size_y; ++y)
             {
-                int x = 0;
-                for(; x <= (pool_size_x - 8); x += 8)
-                {
-                    const q8x8_t data = wrapper::vload(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                                                   (src->info()->strides_in_bytes().y())));
-                    vres              = wrapper::vmax(vres, data);
-                }
-                // Leftover for loop
-                for(; x < pool_size_x; ++x)
+                for(int x = 0; x < pool_size_x; ++x)
                 {
-                    const T data = *(reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int>
-                                                                 (src->info()->strides_in_bytes().y())));
-                    res          = std::max(res, data);
+                    const auto in_ptr = reinterpret_cast<const T *>(in.ptr() + (x - pool_pad_left) * stridex_in_bytes + (y - pool_pad_top) * stridey_in_bytes);
+
+                    const int idx  = x + id.x() * pool_stride_x - pool_pad_left;
+                    const int idy  = y + id.y() * pool_stride_y - pool_pad_top;
+                    const T   data = (idx < 0 || idy < 0 || idx >= src_w || idy >= src_h) ? fill_value : *in_ptr;
+                    res            = std::max(res, data);
                 }
             }
-
-            // Reduce max
-            vres = wrapper::vpmax(vres, vres);
-            vres = wrapper::vpmax(vres, vres);
-            vres = wrapper::vpmax(vres, vres);
-
-            // Get max value
-            res = std::max(res, wrapper::vgetlane(vres, 0));
         }
         // Store result
         res                                 = (src_qinfo != dst_qinfo) ? Qasymm8QuantizationHelper<T>::quantize(Qasymm8QuantizationHelper<T>::dequantize(res, src_qinfo), dst_qinfo) : res;
diff --git a/src/cpu/operators/CpuPool2d.cpp b/src/cpu/operators/CpuPool2d.cpp
index a4ac871d48..eabbd5e0cc 100644
--- a/src/cpu/operators/CpuPool2d.cpp
+++ b/src/cpu/operators/CpuPool2d.cpp
@@ -39,7 +39,6 @@ namespace cpu
 {
 CpuPool2d::CpuPool2d()
     : _pooling_layer_kernel(),
-      _border_handler(),
       _asm_glue(),
       _is_global_pooling_layer(false),
       _data_layout(DataLayout::NCHW),
@@ -86,28 +85,6 @@ void CpuPool2d::configure(ITensorInfo *src, ITensorInfo *dst, const PoolingLayer
         auto k = std::make_unique<kernels::CpuPool2dKernel>();
         k->configure(src, dst, pool_info, indices);
         _pooling_layer_kernel = std::move(k);
-
-        switch(_data_layout)
-        {
-            case DataLayout::NCHW:
-            {
-                // Configure border depending on operation required (quantize border in case of asymmetric data_type)
-                BorderMode border_mode = (!indices && pool_info.pool_type == PoolingType::MAX) ? BorderMode::REPLICATE : BorderMode::CONSTANT;
-                PixelValue zero_value((indices) ? std::numeric_limits<int>::min() : 0.f);
-                if(is_data_type_quantized_asymmetric(src->data_type()) && !pool_info.exclude_padding)
-                {
-                    zero_value = PixelValue(0, src->data_type(), src->quantization_info());
-                }
-                auto b = std::make_unique<NEFillBorderKernel>();
-                b->configure(src, _pooling_layer_kernel->border_size(), border_mode, zero_value);
-                _border_handler = std::move(b);
-                break;
-            }
-            case DataLayout::NHWC:
-                break;
-            default:
-                ARM_COMPUTE_ERROR("Data layout not supported");
-        }
     }
 }
 
@@ -137,14 +114,9 @@ void CpuPool2d::run(ITensorPack &tensors)
         switch(_data_layout)
         {
             case DataLayout::NCHW:
-                // Fill border
-                NEScheduler::get().schedule_op(_border_handler.get(), Window::DimY, _border_handler->window(), tensors);
-
-                // Run pooling layer
                 NEScheduler::get().schedule_op(_pooling_layer_kernel.get(), _is_global_pooling_layer ? Window::DimZ : Window::DimY, _pooling_layer_kernel->window(), tensors);
                 break;
             case DataLayout::NHWC:
-                // Run pooling layer
                 NEScheduler::get().schedule_op(_pooling_layer_kernel.get(), Window::DimX, _pooling_layer_kernel->window(), tensors);
                 break;
             default:
diff --git a/src/cpu/operators/CpuPool2d.h b/src/cpu/operators/CpuPool2d.h
index 471637164f..02c2609a6a 100644
--- a/src/cpu/operators/CpuPool2d.h
+++ b/src/cpu/operators/CpuPool2d.h
@@ -73,7 +73,6 @@ public:
 
 private:
     std::unique_ptr<INEKernel> _pooling_layer_kernel;
-    std::unique_ptr<INEKernel> _border_handler;
     std::unique_ptr<INEKernel> _asm_glue;
 
     bool                             _is_global_pooling_layer;
diff --git a/tests/validation/NEON/PoolingLayer.cpp b/tests/validation/NEON/PoolingLayer.cpp
index a6ba5de11b..77a501582c 100644
--- a/tests/validation/NEON/PoolingLayer.cpp
+++ b/tests/validation/NEON/PoolingLayer.cpp
@@ -193,9 +193,7 @@ TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(RunIndices, NEPoolingLayerIndicesFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerIndicesDatasetFPSmall,
                                                                                                                   framework::dataset::make("DataType",
                                                                                                                           DataType::F16))),
-                                                                                                                  framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })
-
-                                                                                                                 ))
+                                                                                                                  framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f16);