CLInstanceNormalizationLayer NHWC optimisation

* Make changes to split the workload into two kernels. One kernel precomputes
  mean and variance and the second kernel just loads these precomputed values.

* The new approach runs %30 faster than the original code for NHWC workloads
  like 32x192x256.

* Resolves MLCE-337

Change-Id: I8356fcefa2d131ab4dcb32268ce7142421d073e4
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5355
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

6 files changed, 305 insertions, 76 deletions

diff --git a/arm_compute/runtime/CL/functions/CLInstanceNormalizationLayer.h b/arm_compute/runtime/CL/functions/CLInstanceNormalizationLayer.h
index d41f3fedf6..a6e5b1622b 100644
--- a/arm_compute/runtime/CL/functions/CLInstanceNormalizationLayer.h
+++ b/arm_compute/runtime/CL/functions/CLInstanceNormalizationLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,24 +25,44 @@
 #define ARM_COMPUTE_CLINSTANCENORMALIZATIONLAYER_H
 
 #include "arm_compute/core/Error.h"
-#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/IFunction.h"
+
+#include <memory>
 
 namespace arm_compute
 {
 class CLCompileContext;
 class ICLTensor;
 class ITensorInfo;
+class ICLKernel;
+class CLRuntimeContext;
 
 /** Basic function to perform a Instance normalization.
  *
  * This function runs the following kernels:
  * -# @ref CLInstanceNormalizationLayerKernel
  */
-class CLInstanceNormalizationLayer : public ICLSimpleFunction
+class CLInstanceNormalizationLayer : public IFunction
 {
 public:
-    /** Default constructor */
-    CLInstanceNormalizationLayer();
+    /** Constructor
+     *
+     * @param[in] ctx Runtime context to be used by the function
+     */
+    CLInstanceNormalizationLayer(CLRuntimeContext *ctx = nullptr);
+
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLInstanceNormalizationLayer(const CLInstanceNormalizationLayer &) = delete;
+    /** Default move constructor */
+    CLInstanceNormalizationLayer(CLInstanceNormalizationLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLInstanceNormalizationLayer &operator=(const CLInstanceNormalizationLayer &) = delete;
+    /** Default move assignment operator */
+    CLInstanceNormalizationLayer &operator=(CLInstanceNormalizationLayer &&) = default;
+    /** Default destructor */
+    ~CLInstanceNormalizationLayer();
+
     /** Set the input and output tensors.
      *
      * @param[in, out] input               Source tensor. In case of @p output tensor = nullptr this tensor will store the result of the normalization.
@@ -79,6 +99,13 @@ public:
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, float gamma = 1.0f, float beta = 0.0f, float epsilon = 1e-12f, bool use_mixed_precision = true);
+    void run() override;
+
+private:
+    std::unique_ptr<ICLKernel> _inst_norm_kernel; /**< Kernel to run */
+    std::unique_ptr<ICLKernel> _mean_var_kernel;  /**< Kernel to run */
+    CLTensor                   _mean_var_tensor;
+    CLRuntimeContext          *_ctx; /**< Context to use */
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_CLINSTANCENORMALIZATIONLAYER_H */
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index eef204fde9..002a14400f 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -356,6 +356,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "im2col9x9_nhwc", "im2col.cl" },
     { "im2col_generic_nhwc", "im2col.cl" },
     { "instance_normalization", "instance_normalization.cl" },
+    { "compute_mean_var", "instance_normalization.cl" },
     { "l2_normalize_x", "l2_normalize.cl" },
     { "l2_normalize_y", "l2_normalize.cl" },
     { "l2_normalize_z", "l2_normalize.cl" },
diff --git a/src/core/CL/cl_kernels/instance_normalization.cl b/src/core/CL/cl_kernels/instance_normalization.cl
index 480d9cd20c..d2507d94dd 100644
--- a/src/core/CL/cl_kernels/instance_normalization.cl
+++ b/src/core/CL/cl_kernels/instance_normalization.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,14 +23,11 @@
  */
 #include "helpers.h"
 
-#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(INTERNAL_DATA_TYPE) && defined(GAMMA) && defined(BETA) && defined(EPSILON) && defined(DIM_X) && defined(DIM_Y) && defined(DIM_Z)
-/** This function normalizes the input 2D tensor across the first dimension with respect to mean and standard deviation of the same dimension.
+#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DIM_X) && defined(DIM_Y) && defined(DIM_Z)
+/** This function computes the mean and variance of each plane of the input tensor and provides it as output.
  *
  * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
  * @attention Data type should be passed using the -DDATA_TYPE=data_type compile flag, e.g. -DDATA_TYPE=float
- * @attention The scale scalar value applied to the normalized tensor should be passed using the -DGAMMA=value compile flag, e.g. -DGAMMA=1.3
- * @attention The offset scalar value applied to the normalized tensor should be passed using the -DBETA=value compile flag, e.g. -DBETA=2.4
- * @attention Normalization epsilon parameter should be given as a preprocessor argument with -DEPSILON=value. e.g. -DEPSILON=0.001f
  * @attention Dimensions X, Y, and Z should be given as a preprocessor argument with -DDIM_X=value, -DDIM_Y=value, -DDIM_Z=value. e.g. -DDIM_X=6, -DDIM_Y=2, -DDIM_Z=7
  *
  * @param[in]  input_ptr                            Pointer to the first source tensor. Supported data types: F16/F32
@@ -40,6 +37,8 @@
  * @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_stride_w                       Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  input_step_w                         input_stride_w * number of elements along W 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                           (Optional) Pointer to the destination tensor. Supported data types: same as @p input_ptr
  * @param[in]  output_stride_x                      (Optional) Stride of the destination tensor in X dimension (in bytes)
@@ -50,46 +49,40 @@
  * @param[in]  output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
  */
-__kernel void instance_normalization(
-    TENSOR4D_DECLARATION(input)
-#ifndef IN_PLACE
-    ,
-    TENSOR4D_DECLARATION(output)
-#endif /* IN_PLACE */
-)
+__kernel void compute_mean_var(
+    TENSOR4D_DECLARATION(input),
+    TENSOR3D_DECLARATION(output))
 {
-    Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
-#ifndef IN_PLACE
-    Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
-#endif /* IN_PLACE */
-
-    INTERNAL_DATA_TYPE sum    = 0.f;
-    INTERNAL_DATA_TYPE sum_sq = 0.f;
+    Tensor4D in  = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
 
 #if defined(NHWC)
-
     const int ch             = get_global_id(0); // Current channel
-    const int batch          = get_global_id(2); // Current batch
+    const int batch          = get_global_id(1); // Current batch
     const int elements_plane = DIM_Y * DIM_Z;
-
-    for(int i_w = 0; i_w < DIM_Y; ++i_w)
+    float     part_sum       = 0.f;
+    float     part_sum_sq    = 0.f;
+    const int in_offset      = input_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
+    for(int i = 0; i < (DIM_Y * DIM_Z); ++i)
     {
-        for(int i_h = 0; i_h < DIM_Z; ++i_h)
-        {
-            INTERNAL_DATA_TYPE data = (INTERNAL_DATA_TYPE) * ((__global DATA_TYPE *)tensor4D_offset(&in, ch, i_w, i_h, batch));
-            sum += data;
-            sum_sq += data * data;
-        }
+        const float data = *((__global DATA_TYPE *)(input_ptr + in_offset + i * input_stride_y));
+        part_sum += data;
+        part_sum_sq += data * data;
     }
-
+    float    mean                       = (part_sum / elements_plane);
+    float    var                        = (part_sum_sq / elements_plane) - (mean * mean);
+    __global DATA_TYPE *output_address0 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 0, batch);
+    *output_address0                    = mean;
+    __global DATA_TYPE *output_address1 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 1, batch);
+    *output_address1                    = var;
 #else // !defined(NHWC)
     const int ch             = get_global_id(2) % DIM_Z; // Current channel
     const int batch          = get_global_id(2) / DIM_Z; // Current batch
     const int elements_plane = DIM_X * DIM_Y;
 
-    VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE)
+    VEC_DATA_TYPE(float, VEC_SIZE)
     part_sum = 0.f;
-    VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE)
+    VEC_DATA_TYPE(float, VEC_SIZE)
     part_sum_sq = 0.f;
     // Calculate partial sum
     for(int y = 0; y < DIM_Y; ++y)
@@ -98,15 +91,15 @@ __kernel void instance_normalization(
         for(; x <= (DIM_X - VEC_SIZE); x += VEC_SIZE)
         {
             // Load data
-            VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE)
-            data = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)), VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE));
+            VEC_DATA_TYPE(float, VEC_SIZE)
+            data = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)), VEC_DATA_TYPE(float, VEC_SIZE));
             part_sum += data;
             part_sum_sq += data * data;
         }
         // Left-overs loop
         for(; x < DIM_X; ++x)
         {
-            INTERNAL_DATA_TYPE data = (INTERNAL_DATA_TYPE)(*((__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)));
+            float data = (float)(*((__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)));
             part_sum.s0 += data;
             part_sum_sq.s0 += data * data;
         }
@@ -127,29 +120,93 @@ __kernel void instance_normalization(
     part_sum.s0 += part_sum.s1;
     part_sum_sq.s0 += part_sum_sq.s1;
 
-    sum    = (INTERNAL_DATA_TYPE)part_sum.s0;
-    sum_sq = (INTERNAL_DATA_TYPE)part_sum_sq.s0;
+    float sum    = (float)part_sum.s0;
+    float sum_sq = (float)part_sum_sq.s0;
+
+    const float mean = (sum / elements_plane);
+    const float var  = (sum_sq / elements_plane) - (mean * mean);
+
+    __global DATA_TYPE *output_address0 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 0, batch);
+    *output_address0                    = mean;
+    __global DATA_TYPE *output_address1 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 1, batch);
+    *output_address1                    = var;
 
 #endif // defined(NHWC)
+}
+#endif /* defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DIM_X) && defined(DIM_Y) && defined(DIM_Z) */
 
-    const INTERNAL_DATA_TYPE mean   = (sum / elements_plane);
-    const INTERNAL_DATA_TYPE var    = (sum_sq / elements_plane) - (mean * mean);
-    const INTERNAL_DATA_TYPE multip = GAMMA / sqrt(var + EPSILON);
+#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(INTERNAL_DATA_TYPE) && defined(GAMMA) && defined(BETA) && defined(EPSILON) && defined(DIM_X) && defined(DIM_Y) && defined(DIM_Z)
+/** This function normalizes the input 2D tensor across the first dimension with respect to mean and standard deviation of the same dimension.
+ *
+ * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @attention Data type should be passed using the -DDATA_TYPE=data_type compile flag, e.g. -DDATA_TYPE=float
+ * @attention The scale scalar value applied to the normalized tensor should be passed using the -DGAMMA=value compile flag, e.g. -DGAMMA=1.3
+ * @attention The offset scalar value applied to the normalized tensor should be passed using the -DBETA=value compile flag, e.g. -DBETA=2.4
+ * @attention Normalization epsilon parameter should be given as a preprocessor argument with -DEPSILON=value. e.g. -DEPSILON=0.001f
+ * @attention Dimensions X, Y, and Z should be given as a preprocessor argument with -DDIM_X=value, -DDIM_Y=value, -DDIM_Z=value. e.g. -DDIM_X=6, -DDIM_Y=2, -DDIM_Z=7
+ *
+ * @param[in]  input_ptr                            Pointer to the first source tensor. Supported data types: 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                           (Optional) Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      (Optional) Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      (Optional) Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      (Optional) Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
+ */
+__kernel void instance_normalization(
+    TENSOR4D_DECLARATION(input),
+    TENSOR3D_DECLARATION(mean_var)
+#ifndef IN_PLACE
+    ,
+    TENSOR4D_DECLARATION(output)
+#endif /* IN_PLACE */
+)
+{
+    Tensor4D in       = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
+    Tensor3D mean_var = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(mean_var);
+#ifndef IN_PLACE
+    Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+#endif /* IN_PLACE */
 
 #if defined(NHWC)
+    const int ch    = get_global_id(0); // Current channel
+    const int batch = get_global_id(2); // Current batch
+#else                                   /* defined(NHWC) */
+    const int ch    = get_global_id(2) % DIM_Z; // Current channel
+    const int batch = get_global_id(2) / DIM_Z; // Current batch
+#endif                                  /* defined(NHWC) */
 
-    for(int i_w = 0; i_w < DIM_Y; ++i_w)
+    const __global DATA_TYPE *mean_ptr                   = (__global DATA_TYPE *)tensor3D_offset(&mean_var, ch, 0, batch);
+    const __global DATA_TYPE *var_ptr                    = (__global DATA_TYPE *)tensor3D_offset(&mean_var, ch, 1, batch);
+    const INTERNAL_DATA_TYPE                      mean   = (INTERNAL_DATA_TYPE) * mean_ptr;
+    const INTERNAL_DATA_TYPE                      var    = (INTERNAL_DATA_TYPE) * var_ptr;
+    const INTERNAL_DATA_TYPE                      multip = GAMMA / sqrt(var + EPSILON);
+    const INTERNAL_DATA_TYPE                      beta   = (INTERNAL_DATA_TYPE)BETA;
+
+#if defined(NHWC)
+    const int in_offset = input_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
+#ifndef IN_PLACE
+    const int out_offset = output_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
+#endif /* IN_PLACE */
+
+    for(int i = 0; i < (DIM_Y * DIM_Z); ++i)
     {
-        for(int i_h = 0; i_h < DIM_Z; ++i_h)
-        {
-            __global DATA_TYPE *input_address = (__global DATA_TYPE *)tensor4D_offset(&in, ch, i_w, i_h, batch);
+        __global DATA_TYPE *input_address = (__global DATA_TYPE *)(input_ptr + in_offset + i * input_stride_y);
 #ifdef IN_PLACE
-            __global DATA_TYPE *output_address = input_address;
+        __global DATA_TYPE *output_address = input_address;
 #else  /* !IN_PLACE */
-            __global DATA_TYPE *output_address = (__global DATA_TYPE *)tensor4D_offset(&out, ch, i_w, i_h, batch);
+        __global DATA_TYPE *output_address = (__global DATA_TYPE *)(output_ptr + out_offset + i * output_stride_y);
 #endif /* IN_PLACE */
-            *(output_address) = (*(input_address) - mean) * multip + (INTERNAL_DATA_TYPE)BETA;
-        }
+        *(output_address) = (*(input_address) - mean) * multip + beta;
     }
 
 #else // !defined(NHWC)
diff --git a/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp
index 50c4e24c33..80a42cc3f5 100644
--- a/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp
@@ -32,7 +32,6 @@
 #include "src/core/CL/CLValidate.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
-
 #include "support/StringSupport.h"
 
 namespace arm_compute
@@ -54,25 +53,108 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
 
     return Status{};
 }
+
+Status validate_arguments_meanvar(const ITensorInfo *input, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(input, DataType::F16, DataType::F32);
+
+    if(output != nullptr && output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_channels() != output->num_channels(), "Input and output have different number of channels");
+    }
+
+    return Status{};
+}
 } // namespace
 
-CLInstanceNormalizationLayerKernel::CLInstanceNormalizationLayerKernel()
-    : _input(nullptr), _output(nullptr), _run_in_place(false)
+CLComputeMeanVariance::CLComputeMeanVariance()
+    : _input(nullptr), _output(nullptr)
+{
+}
+
+void CLComputeMeanVariance::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+    auto padding_info = get_padding_info({ input, output });
+
+    _input  = input;
+    _output = output == nullptr ? input : output;
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_meanvar(_input->info(), _output->info()));
+    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
+    build_opts.add_option("-DDIM_X=" + support::cpp11::to_string(input->info()->dimension(0)));
+    build_opts.add_option("-DDIM_Y=" + support::cpp11::to_string(input->info()->dimension(1)));
+    build_opts.add_option("-DDIM_Z=" + support::cpp11::to_string(input->info()->dimension(2)));
+    build_opts.add_option_if(_input->info()->data_layout() == DataLayout::NHWC, "-DNHWC");
+    // Create kernel
+    _kernel = create_kernel(compile_context, "compute_mean_var", build_opts.options());
+
+    // We handle the planes manually
+    Window             win           = calculate_max_window(*(input->info()), Steps(1));
+    const auto         data_layout   = input->info()->data_layout();
+    const unsigned int channel_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+    const unsigned int batches_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
+    const unsigned int input_channel = input->info()->dimension(channel_idx);
+    const unsigned int input_batches = input->info()->dimension(batches_idx);
+    const TensorShape  out_shape(input_channel, 2u, input_batches);
+
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), out_shape, 1, input->info()->data_type());
+
+    ICLKernel::configure_internal(win);
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
+}
+
+Status CLComputeMeanVariance::validate(const ITensorInfo *input, const ITensorInfo *output)
 {
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_meanvar(input, output));
+    return Status{};
 }
 
-void CLInstanceNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *output, const InstanceNormalizationLayerKernelInfo &info)
+void CLComputeMeanVariance::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    Window collapsed_window = window.collapse(window, Window::DimZ);
+
+    // We will process the planes together
+    if(_input->info()->data_layout() == DataLayout::NCHW)
+    {
+        collapsed_window.set(Window::DimX, Window::Dimension(0, 1, 1));
+        collapsed_window.set(Window::DimY, Window::Dimension(0, 1, 1));
+    }
+    else
+    {
+        collapsed_window.set(Window::DimZ, Window::Dimension(0, 1, 1));
+        collapsed_window.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(3), 1));
+    }
+    unsigned int idx = 0;
+    add_4D_tensor_argument(idx, _input, collapsed_window);
+    add_3D_tensor_argument(idx, _output, collapsed_window);
+
+    enqueue(queue, *this, collapsed_window, lws_hint());
+}
+
+CLInstanceNormalizationLayerKernel::CLInstanceNormalizationLayerKernel()
+    : _input(nullptr), _output(nullptr), _mean(nullptr), _run_in_place(false)
 {
-    configure(CLKernelLibrary::get().get_compile_context(), input, output, info);
 }
 
-void CLInstanceNormalizationLayerKernel::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, const InstanceNormalizationLayerKernelInfo &info)
+void CLInstanceNormalizationLayerKernel::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *mean_var, ICLTensor *output, const InstanceNormalizationLayerKernelInfo &info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input);
     auto padding_info = get_padding_info({ input, output });
 
     _input  = input;
     _output = output == nullptr ? input : output;
+    _mean   = mean_var;
 
     _run_in_place = (output == nullptr) || (output == input);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(_input->info(), _output->info(), info));
@@ -132,6 +214,8 @@ void CLInstanceNormalizationLayerKernel::run(const Window &window, cl::CommandQu
 
     unsigned int idx = 0;
     add_4D_tensor_argument(idx, _input, collapsed_window);
+    add_3D_tensor_argument(idx, _mean, collapsed_window);
+
     if(!_run_in_place)
     {
         add_4D_tensor_argument(idx, _output, collapsed_window);
diff --git a/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h b/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h
index d4444f0b20..33a3ff97c3 100644
--- a/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h
+++ b/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -52,21 +52,14 @@ public:
 
     /** Set the input and output tensors.
      *
-     * @param[in, out] input  Source tensor. Data types supported: F16/F32. Data layout supported: NCHW, NHWC
-     *                        In case of @p output tensor = nullptr this tensor will store the result of the normalization.
-     * @param[out]     output Destination tensor. Data types and data layouts supported: same as @p input.
-     * @param[in]      info   Kernel meta-data descriptor
-     */
-    void configure(ICLTensor *input, ICLTensor *output, const InstanceNormalizationLayerKernelInfo &info);
-    /** Set the input and output tensors.
-     *
      * @param[in]      compile_context The compile context to be used.
      * @param[in, out] input           Source tensor. Data types supported: F16/F32. Data layout supported: NCHW, NHWC
      *                                 In case of @p output tensor = nullptr this tensor will store the result of the normalization.
+     * @param[in]      mean_var        Tensor containing the precomputed mean and variance values. Data types supported: F32.
      * @param[out]     output          Destination tensor. Data types and data layouts supported: same as @p input.
      * @param[in]      info            Kernel meta-data descriptor
      */
-    void configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, const InstanceNormalizationLayerKernelInfo &info);
+    void configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *mean_var, ICLTensor *output, const InstanceNormalizationLayerKernelInfo &info);
 
     /** Static function to check if given info will lead to a valid configuration of @ref CLInstanceNormalizationLayer.
      *
@@ -84,7 +77,51 @@ public:
 private:
     ICLTensor *_input;
     ICLTensor *_output;
+    ICLTensor *_mean;
     bool       _run_in_place;
 };
+
+/** Interface for compute Mean and Variance per channel */
+class CLComputeMeanVariance : public ICLKernel
+{
+public:
+    /** Constructor */
+    CLComputeMeanVariance();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLComputeMeanVariance(const CLComputeMeanVariance &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLComputeMeanVariance &operator=(const CLComputeMeanVariance &) = delete;
+    /** Default Move Constructor. */
+    CLComputeMeanVariance(CLComputeMeanVariance &&) = default;
+    /** Default move assignment operator */
+    CLComputeMeanVariance &operator=(CLComputeMeanVariance &&) = default;
+    /** Default destructor */
+    ~CLComputeMeanVariance() = default;
+
+    /** Set the input and output tensors.
+     *
+     * @param[in]      compile_context The compile context to be used.
+     * @param[in, out] input           Source tensor. Data types supported: F16/F32. Data layout supported: NCHW, NHWC
+     *                                 In case of @p output tensor = nullptr this tensor will store the result of the normalization.
+     * @param[out]     output          Destination tensor. Data types and data layouts supported: same as @p input.
+     */
+    void configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref CLInstanceNormalizationLayer.
+     *
+     * @param[in] input  Source tensor info. Data types supported: F16/F32. Data layout supported: NHWC, NCHW
+     * @param[in] output Destination tensor info. Data types and data layouts supported: same as @p input.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    ICLTensor *_input;
+    ICLTensor *_output;
+};
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_CLINSTANCENORMALIZATIONLAYERKERNEL_H */
diff --git a/src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp b/src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp
index 9bc060e6ca..f2406d68f4 100644
--- a/src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp
+++ b/src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,13 +23,24 @@
  */
 #include "arm_compute/runtime/CL/functions/CLInstanceNormalizationLayer.h"
 
+#include "arm_compute/core/Error.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLHelpers.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+#include "src/core/CL/ICLKernel.h"
 #include "src/core/CL/kernels/CLFillBorderKernel.h"
 #include "src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h"
 
 namespace arm_compute
 {
-CLInstanceNormalizationLayer::CLInstanceNormalizationLayer()
+CLInstanceNormalizationLayer::CLInstanceNormalizationLayer(CLRuntimeContext *ctx) // NOLINT
+    : _inst_norm_kernel(),
+      _mean_var_kernel(),
+      _mean_var_tensor(),
+      _ctx(ctx)
+{
+}
+CLInstanceNormalizationLayer::~CLInstanceNormalizationLayer()
 {
 }
 
@@ -40,13 +51,25 @@ void CLInstanceNormalizationLayer::configure(ICLTensor *input, ICLTensor *output
 
 void CLInstanceNormalizationLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, float gamma, float beta, float epsilon, bool use_mixed_precision)
 {
-    auto k = std::make_unique<CLInstanceNormalizationLayerKernel>();
-    k->configure(compile_context, input, output, InstanceNormalizationLayerKernelInfo(gamma, beta, epsilon, use_mixed_precision));
-    _kernel = std::move(k);
+    auto w = std::make_unique<CLComputeMeanVariance>();
+    w->configure(compile_context, input, &_mean_var_tensor);
+    _mean_var_kernel = std::move(w);
+    auto k           = std::make_unique<CLInstanceNormalizationLayerKernel>();
+    k->configure(compile_context, input, &_mean_var_tensor, output, InstanceNormalizationLayerKernelInfo(gamma, beta, epsilon, use_mixed_precision));
+    _inst_norm_kernel = std::move(k);
+    _mean_var_tensor.allocator()->allocate();
 }
 
 Status CLInstanceNormalizationLayer::validate(const ITensorInfo *input, const ITensorInfo *output, float gamma, float beta, float epsilon, bool use_mixed_precision)
 {
     return CLInstanceNormalizationLayerKernel::validate(input, output, InstanceNormalizationLayerKernelInfo(gamma, beta, epsilon, use_mixed_precision));
 }
-} // namespace arm_compute
\ No newline at end of file
+
+void CLInstanceNormalizationLayer::run()
+{
+    ARM_COMPUTE_ERROR_ON_MSG(!_inst_norm_kernel, "The child class didn't set the CL kernel or function isn't configured");
+    schedule_kernel_on_ctx(_ctx, _mean_var_kernel.get());
+    schedule_kernel_on_ctx(_ctx, _inst_norm_kernel.get());
+}
+
+} // namespace arm_compute