COMPMID-477 - Optimized CLNormalizationLayer

CLPixelWiseMultiplication has been removed within the function

Change-Id: Ibe7edd7921d5cef6ff68fdeeca89771129a8eaea
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/84459
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

5 files changed, 82 insertions, 114 deletions

diff --git a/arm_compute/core/CL/kernels/CLNormalizationLayerKernel.h b/arm_compute/core/CL/kernels/CLNormalizationLayerKernel.h
index 5eedc31486..5f8c9c9d07 100644
--- a/arm_compute/core/CL/kernels/CLNormalizationLayerKernel.h
+++ b/arm_compute/core/CL/kernels/CLNormalizationLayerKernel.h
@@ -48,14 +48,12 @@ public:
 
     /** Set the input and output tensors.
      *
-     * @param[in]  input         Source tensor. 3 lower dims represent a single input with dimensions [width, height, IFM],
-     *                           and an optional 4th dimension for batch of inputs. Data types supported: QS8/QS16/F16/F32.
-     * @param[in]  squared_input Source with each element has been squared. 3 lower dims represent a single input with dimensions [width, height, IFM].
-     *                           Data types supported: same as @p input.
-     * @param[out] output        Destination tensor. Output will have the same number of dimensions as input. Data types supported: same as @p input.
-     * @param[in]  norm_info     Normalization layer information like the normalization type, normalization size and other parameters.
+     * @param[in]  input     Source tensor. 3 lower dims represent a single input with dimensions [width, height, IFM],
+     *                       and an optional 4th dimension for batch of inputs. Data types supported: QS8/QS16/F16/F32.
+     * @param[out] output    Destination tensor. Output will have the same number of dimensions as input. Data types supported: same as @p input.
+     * @param[in]  norm_info Normalization layer information like the normalization type, normalization size and other parameters.
      */
-    void configure(const ICLTensor *input, const ICLTensor *squared_input, ICLTensor *output, NormalizationLayerInfo norm_info);
+    void configure(const ICLTensor *input, ICLTensor *output, NormalizationLayerInfo norm_info);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
@@ -63,7 +61,6 @@ public:
 
 private:
     const ICLTensor *_input;
-    const ICLTensor *_squared_input;
     ICLTensor       *_output;
     BorderSize       _border_size;
     bool             _is_in_map;
diff --git a/arm_compute/runtime/CL/functions/CLNormalizationLayer.h b/arm_compute/runtime/CL/functions/CLNormalizationLayer.h
index a4dae85c1d..0818cec2e5 100644
--- a/arm_compute/runtime/CL/functions/CLNormalizationLayer.h
+++ b/arm_compute/runtime/CL/functions/CLNormalizationLayer.h
@@ -39,7 +39,6 @@ class ICLTensor;
 
 /** Basic function to simulate a normalization layer. This function calls the following CL kernels:
  *
- * -# @ref CLPixelWiseMultiplicationKernel
  * -# @ref CLFillBorderKernel
  * -# @ref CLNormalizationLayerKernel
  *
@@ -51,21 +50,19 @@ public:
     CLNormalizationLayer();
     /** Set the input and output tensors.
      *
-     * @param[in]  input     Source tensor. 3 lower dims represent a single input with dimensions [width, height, IFM],
-     *                       and an optional 4th dimension for batch of inputs. Data types supported: F16, F32. Number of channels must be 1.
-     * @param[out] output    Destination tensor. Dimensions, data type and number of channels must match the input ones.
-     * @param[in]  norm_info Normalization layer information like the normalization type, normalization size and other parameters.
+     * @param[in, out] input     Source tensor. 3 lower dims represent a single input with dimensions [width, height, IFM],
+     *                           and an optional 4th dimension for batch of inputs. Data types supported: F16/F32 (Written to by the border handler)
+     * @param[out]     output    Destination tensor. Dimensions, data type and number of channels must match the input ones.
+     * @param[in]      norm_info Normalization layer information like the normalization type, normalization size and other parameters.
      */
-    void configure(const ICLTensor *input, ICLTensor *output, NormalizationLayerInfo norm_info);
+    void configure(ICLTensor *input, ICLTensor *output, NormalizationLayerInfo norm_info);
 
     // Inherited methods overridden:
     void run() override;
 
 private:
-    CLTensor                        _squared_input;   /**< The intermediate buffer which stores results of squaring input*/
-    CLNormalizationLayerKernel      _norm_kernel;     /**< Normalization layer kernel to run */
-    CLPixelWiseMultiplicationKernel _multiply_kernel; /**< Pixel multiplication kernel to run */
-    CLFillBorderKernel              _border_handler;  /**< Kernel to handle  borders */
+    CLNormalizationLayerKernel _norm_kernel;    /**< Normalization layer kernel to run */
+    CLFillBorderKernel         _border_handler; /**< Kernel to handle  borders */
 };
 }
 #endif /* __ARM_COMPUTE_CLNORMALIZATIONLAYER_H__ */
diff --git a/src/core/CL/cl_kernels/normalization_layer.cl b/src/core/CL/cl_kernels/normalization_layer.cl
index e2a5c4079a..4e65560b95 100644
--- a/src/core/CL/cl_kernels/normalization_layer.cl
+++ b/src/core/CL/cl_kernels/normalization_layer.cl
@@ -54,43 +54,33 @@
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16
+ * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5
+ * @note The number of slices should be given as a preprocessor argument using -DNUM_SLICES=size. e.g. -DNUM_SLICES=192
  * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3
  * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA
  *
- * @param[in]  input_ptr                                   Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32
- * @param[in]  input_stride_x                              Stride of the first source tensor in X dimension (in bytes)
- * @param[in]  input_step_x                                input_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  input_stride_y                              Stride of the first source tensor in Y dimension (in bytes)
- * @param[in]  input_step_y                                input_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  input_stride_z                              Stride of the first source tensor in Z dimension (in bytes)
- * @param[in]  input_step_z                                input_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  input_offset_first_element_in_bytes         The offset of the first element in the first source tensor
- * @param[in]  squared_input_ptr                           Pointer to the second source tensor. Supported data types: same as @p input_ptr
- * @param[in]  squared_input_stride_x                      Stride of the second source tensor in X dimension (in bytes)
- * @param[in]  squared_input_step_x                        input_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  squared_input_stride_y                      Stride of the second source tensor in Y dimension (in bytes)
- * @param[in]  squared_input_step_y                        input_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  squared_input_stride_z                      Stride of the second source tensor in Z dimension (in bytes)
- * @param[in]  squared_input_step_z                        input_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  squared_input_offset_first_element_in_bytes The offset of the second element in the second source tensor
- * @param[out] output_ptr                                  Pointer to the destination tensor. Supported data types: same as @p input_ptr
- * @param[in]  output_stride_x                             Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  output_step_x                               output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  output_stride_y                             Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  output_step_y                               output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  output_stride_z                             Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  output_step_z                               output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  output_offset_first_element_in_bytes        The offset of the first element in the destination tensor
- * @param[in]  radius                                      Number of elements on the right or left side to normalize across
+ * @param[in]  input_ptr                            Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32
+ * @param[in]  input_stride_x                       Stride of the first source tensor in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the first source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the first source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the first source tensor
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination tensor
  */
 __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input),
-                                            TENSOR3D_DECLARATION(squared_input),
-                                            TENSOR3D_DECLARATION(output),
-                                            uint radius)
+                                            TENSOR3D_DECLARATION(output))
 {
-    Tensor3D in         = CONVERT_TO_TENSOR3D_STRUCT(input);
-    Tensor3D squared_in = CONVERT_TO_TENSOR3D_STRUCT(squared_input);
-    Tensor3D out        = CONVERT_TO_TENSOR3D_STRUCT(output);
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
 
     VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
     acc = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))0;
@@ -101,15 +91,16 @@ __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input),
     const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
     kappa_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(KAPPA);
 
-    const int num_of_slices = get_global_size(2);
     const int current_slice = get_global_id(2);
 
-    const int left_slice  = max(current_slice - (int)radius, (int)0);
-    const int right_slice = min(current_slice + (int)radius, (int)(num_of_slices - 1));
+    const int left_slice  = max(current_slice - (int)RADIUS, (int)0);
+    const int right_slice = min(current_slice + (int)RADIUS, (int)(NUM_SLICES - 1));
 
     for(int i = left_slice; i <= right_slice; i++)
     {
-        acc = ADD_OP(acc, LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&squared_in, 0, 0, i - current_slice)));
+        VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+        values = LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&in, 0, 0, i - current_slice));
+        acc    = ADD_OP(acc, MUL_OP(values, values));
     }
 
     acc = ADD_OP(MUL_OP(acc, coeff_v), kappa_v);
@@ -125,43 +116,32 @@ __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input),
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16
+ * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5
  * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3
  * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA
  *
- * @param[in]  input_ptr                                   Pointer to the first source tensor. Supported data types: QS8/F16/F32
- * @param[in]  input_stride_x                              Stride of the first source tensor in X dimension (in bytes)
- * @param[in]  input_step_x                                input_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  input_stride_y                              Stride of the first source tensor in Y dimension (in bytes)
- * @param[in]  input_step_y                                input_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  input_stride_z                              Stride of the first source tensor in Z dimension (in bytes)
- * @param[in]  input_step_z                                input_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  input_offset_first_element_in_bytes         The offset of the first element in the first source tensor
- * @param[in]  squared_input_ptr                           Pointer to the second source tensor. Supported data types: same as @p input_ptr
- * @param[in]  squared_input_stride_x                      Stride of the second source tensor in X dimension (in bytes)
- * @param[in]  squared_input_step_x                        input_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  squared_input_stride_y                      Stride of the second source tensor in Y dimension (in bytes)
- * @param[in]  squared_input_step_y                        input_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  squared_input_stride_z                      Stride of the second source tensor in Z dimension (in bytes)
- * @param[in]  squared_input_step_z                        input_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  squared_input_offset_first_element_in_bytes The offset of the second element in the second source tensor
- * @param[out] output_ptr                                  Pointer to the destination tensor. Supported data types: same as @p input_ptr
- * @param[in]  output_stride_x                             Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  output_step_x                               output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  output_stride_y                             Stride of the first destination tensor in Y dimension (in bytes)
- * @param[in]  output_step_y                               output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  output_stride_z                             Stride of the first source tensor in Z dimension (in bytes)
- * @param[in]  output_step_z                               output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  output_offset_first_element_in_bytes        The offset of the first element in the destination tensor
- * @param[in]  radius                                      Number of elements on the right or left side to normalize across
+ * @param[in]  input_ptr                            Pointer to the first source tensor. Supported data types: QS8/F16/F32
+ * @param[in]  input_stride_x                       Stride of the first source tensor in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the first source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the first source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the first source tensor
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the first destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the first source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination tensor
  */
 __kernel void normalization_layer_in_map_1D(TENSOR3D_DECLARATION(input),
-                                            TENSOR3D_DECLARATION(squared_input),
-                                            TENSOR3D_DECLARATION(output),
-                                            uint radius)
+                                            TENSOR3D_DECLARATION(output))
 {
-    Tensor3D in         = CONVERT_TO_TENSOR3D_STRUCT(input);
-    Tensor3D squared_in = CONVERT_TO_TENSOR3D_STRUCT(squared_input);
-    Tensor3D out        = CONVERT_TO_TENSOR3D_STRUCT(output);
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
 
     VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
     acc = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))0;
@@ -174,12 +154,14 @@ __kernel void normalization_layer_in_map_1D(TENSOR3D_DECLARATION(input),
 
     const int current_pos = get_global_id(0) << 2;
 
-    const int left_pos  = max(current_pos - (int)radius, -3);
-    const int right_pos = min(current_pos + (int)radius, (int)((get_global_size(0) << 2) + 3 - 1));
+    const int left_pos  = max(current_pos - (int)RADIUS, -3);
+    const int right_pos = min(current_pos + (int)RADIUS, (int)((get_global_size(0) << 2) + 3 - 1));
 
     for(int i = left_pos; i <= right_pos; i += 1)
     {
-        acc = ADD_OP(acc, LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&squared_in, i - current_pos, 0, 0)));
+        VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+        values = LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&in, i - current_pos, 0, 0));
+        acc    = ADD_OP(acc, MUL_OP(values, values));
     }
 
     acc = ADD_OP(MUL_OP(acc, coeff_v), kappa_v);
diff --git a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
index b382e9d510..a74473980b 100644
--- a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
@@ -36,7 +36,7 @@
 using namespace arm_compute;
 
 CLNormalizationLayerKernel::CLNormalizationLayerKernel()
-    : _input(nullptr), _squared_input(nullptr), _output(nullptr), _border_size(0), _is_in_map(false)
+    : _input(nullptr), _output(nullptr), _border_size(0), _is_in_map(false)
 {
 }
 
@@ -45,7 +45,7 @@ BorderSize CLNormalizationLayerKernel::border_size() const
     return _border_size;
 }
 
-void CLNormalizationLayerKernel::configure(const ICLTensor *input, const ICLTensor *squared_input, ICLTensor *output, NormalizationLayerInfo norm_info)
+void CLNormalizationLayerKernel::configure(const ICLTensor *input, ICLTensor *output, NormalizationLayerInfo norm_info)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
@@ -53,21 +53,20 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, const ICLTens
     // Output tensor auto initialization if not yet initialized
     auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position());
 
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, squared_input, output);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, squared_input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
     ARM_COMPUTE_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd");
     ARM_COMPUTE_ERROR_ON_MSG(norm_info.type() == NormType::IN_MAP_2D, "2D In-Map Normalization not implemented");
     if(is_data_type_fixed_point(input->info()->data_type()))
     {
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, squared_input, output);
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
         ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.beta(), input);
         ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.kappa(), input);
         ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.scale_coeff(), input);
     }
 
-    _input         = input;
-    _squared_input = squared_input;
-    _output        = output;
+    _input  = input;
+    _output = output;
 
     _is_in_map                      = (norm_info.type() != NormType::CROSS_MAP);
     const unsigned int border_width = _is_in_map ? std::min(norm_info.norm_size() / 2, 3U) : 0;
@@ -87,23 +86,20 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, const ICLTens
     build_opts.emplace(("-DBETA=" + float_to_string_with_full_precision(norm_info.beta())));
     build_opts.emplace(("-DKAPPA=" + float_to_string_with_full_precision(norm_info.kappa())));
     build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
+    build_opts.emplace(("-DRADIUS=" + support::cpp11::to_string(norm_info.norm_size() / 2)));
+    build_opts.emplace(("-DNUM_SLICES=" + support::cpp11::to_string(input->info()->dimension(2))));
 
     // Create kernel
     std::string kernel_name = (norm_info.type() == NormType::IN_MAP_1D) ? "normalization_layer_in_map_1D" : "normalization_layer_cross_map";
     _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
 
-    // Set kernel static arguments
-    unsigned int idx = 3 * num_arguments_per_3D_tensor(); // Skip the input and output parameters
-    _kernel.setArg<cl_uint>(idx++, norm_info.norm_size() / 2);
-
     // Configure kernel window
     Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
 
     AccessWindowHorizontal input_access(input->info(), -_border_size.left, num_elems_read_per_iteration);
-    AccessWindowHorizontal squared_input_access(squared_input->info(), -_border_size.left, num_elems_read_per_iteration);
     AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
 
-    update_window_and_padding(win, input_access, squared_input_access, output_access);
+    update_window_and_padding(win, input_access, output_access);
 
     output_access.set_valid_region(win, input->info()->valid_region());
 
@@ -123,7 +119,6 @@ void CLNormalizationLayerKernel::run(const Window &window, cl::CommandQueue &que
     {
         unsigned int idx = 0;
         add_3D_tensor_argument(idx, _input, slice);
-        add_3D_tensor_argument(idx, _squared_input, slice);
         add_3D_tensor_argument(idx, _output, slice);
         enqueue(queue, *this, slice);
     }
diff --git a/src/runtime/CL/functions/CLNormalizationLayer.cpp b/src/runtime/CL/functions/CLNormalizationLayer.cpp
index 69cef334e8..f4bd49406c 100644
--- a/src/runtime/CL/functions/CLNormalizationLayer.cpp
+++ b/src/runtime/CL/functions/CLNormalizationLayer.cpp
@@ -33,29 +33,26 @@
 using namespace arm_compute;
 
 CLNormalizationLayer::CLNormalizationLayer()
-    : _squared_input(), _norm_kernel(), _multiply_kernel(), _border_handler()
+    : _norm_kernel(), _border_handler()
 {
 }
 
-void CLNormalizationLayer::configure(const ICLTensor *input, ICLTensor *output, NormalizationLayerInfo norm_info)
+void CLNormalizationLayer::configure(ICLTensor *input, ICLTensor *output, NormalizationLayerInfo norm_info)
 {
     ARM_COMPUTE_ERROR_ON(input == nullptr);
 
-    TensorInfo tensor_info(input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position());
-    _squared_input.allocator()->init(tensor_info);
+    // Configure normalization kernel
+    _norm_kernel.configure(input, output, norm_info);
 
-    _norm_kernel.configure(input, &_squared_input, output, norm_info);
-    _multiply_kernel.configure(input, input, &_squared_input, 1.0f, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
     // Fill the border by 3 elements since we need vload4 in the IN_MAP normalization kernel
-    _border_handler.configure(&_squared_input, _norm_kernel.border_size(), BorderMode::CONSTANT, PixelValue(0));
-
-    // Allocate intermediate buffers
-    _squared_input.allocator()->allocate();
+    _border_handler.configure(input, _norm_kernel.border_size(), BorderMode::CONSTANT, PixelValue(0));
 }
 
 void CLNormalizationLayer::run()
 {
-    CLScheduler::get().enqueue(_multiply_kernel, false);
+    // Run border handler
     CLScheduler::get().enqueue(_border_handler, false);
-    CLScheduler::get().enqueue(_norm_kernel, false);
+
+    // Run normalization kernel
+    CLScheduler::get().enqueue(_norm_kernel);
 }