COMPMID-2342: Add layer normalization support in CLLSTMLayer

Change-Id: I25d974aa94e69c5f79a0bd99d5869a351d6d954d
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1324
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

6 files changed, 426 insertions, 107 deletions

diff --git a/arm_compute/runtime/CL/functions/CLLSTMLayer.h b/arm_compute/runtime/CL/functions/CLLSTMLayer.h
index 3add152878..44b3baf81b 100644
--- a/arm_compute/runtime/CL/functions/CLLSTMLayer.h
+++ b/arm_compute/runtime/CL/functions/CLLSTMLayer.h
@@ -39,6 +39,7 @@
 #include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h"
 #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h"
 #include "arm_compute/runtime/CL/functions/CLGEMM.h"
+#include "arm_compute/runtime/CL/functions/CLMeanStdDevNormalizationLayer.h"
 #include "arm_compute/runtime/IMemoryManager.h"
 #include "arm_compute/runtime/common/LSTMParams.h"
 
@@ -76,17 +77,22 @@ public:
      * @param[out] output                      Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].
      *                                         Data types supported: Same as @p input.
      * @param[in]  lstm_params                 (Optional) Weights tensors used in peephole optimization:
-     *                                         input_to_input_weights       2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
-     *                                         recurrent_to_input_weights   2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
-     *                                         cell_to_input_weights        1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
-     *                                         cell_to_forget_weights       1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
-     *                                         cell_to_output_weights       1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
-     *                                         input_gate_bias              1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
-     *                                         projection_weights           2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
-     *                                         projection_bias              1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
+     *                                         input_to_input_weights         2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
+     *                                         recurrent_to_input_weights     2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     *                                         cell_to_input_weights          1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
+     *                                         cell_to_forget_weights         1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         cell_to_output_weights         1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         input_gate_bias                1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
+     *                                         projection_weights             2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     *                                         projection_bias                1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
+     *                                         input_layer_norm_coefficients  1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         forget_layer_norm_coefficients 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         cell_layer_norm_coefficients   1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         output_layer_norm_coefficients 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
      * @param[in]  activation_info             Contains activation information described in @ref ActivationLayerInfo.
-     * @param[in]  cell_threshold              The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. If set to 0.0 then clipping is disabled.
-     * @param[in]  projection_threshold        The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled.
+     * @param[in]  cell_threshold              The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. If set to 0.0f then clipping is disabled.
+     * @param[in]  projection_threshold        The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip].
+     *                                         If set to 0.0f then clipping is disabled.
      */
     void configure(const ICLTensor *input,
                    const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
@@ -98,35 +104,40 @@ public:
 
     /** Static function to check if given info will lead to a valid configuration of @ref CLLSTMLayer
      *
-     * @param[in] input                       Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32.
-     * @param[in] input_to_forget_weights     2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
-     * @param[in] input_to_cell_weights       2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
-     * @param[in] input_to_output_weights     2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
-     * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
-     * @param[in] recurrent_to_cell_weights   2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
-     * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
-     * @param[in] forget_gate_bias            1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
-     * @param[in] cell_bias                   1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
-     * @param[in] output_gate_bias            1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
-     * @param[in] output_state_in             2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
-     * @param[in] cell_state_in               2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
-     * @param[in] scratch_buffer              2D tensor with dimensions [num_units * 4, batch_size] with CIFG or [num_units * 3, batch_size] without CIGF. Data type supported: Same as @p input.
-     * @param[in] output_state_out            2D weights tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
-     * @param[in] cell_state_out              2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
-     * @param[in] output                      Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].
-     *                                        Data types supported: Same as @p input.
-     * @param[in] lstm_params                 (Optional) Weights tensors used in peephole optimization:
-     *                                        input_to_input_weights       2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
-     *                                        recurrent_to_input_weights   2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
-     *                                        cell_to_input_weights        1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
-     *                                        cell_to_forget_weights       1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
-     *                                        cell_to_output_weights       1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
-     *                                        input_gate_bias              1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
-     *                                        projection_weights           2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
-     *                                        projection_bias              1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
+     * @param[in] input                       Source tensor info. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: F16/F32.
+     * @param[in] input_to_forget_weights     2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
+     * @param[in] input_to_cell_weights       2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
+     * @param[in] input_to_output_weights     2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
+     * @param[in] recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     * @param[in] recurrent_to_cell_weights   2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     * @param[in] recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     * @param[in] forget_gate_bias            1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     * @param[in] cell_bias                   1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     * @param[in] output_gate_bias            1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     * @param[in] output_state_in             2D weights tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
+     * @param[in] cell_state_in               2D tensor info with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
+     * @param[in] scratch_buffer              2D tensor info with dimensions [num_units * 4, batch_size] with CIFG or [num_units * 3, batch_size] without CIGF.
+     *                                        Data type supported: Same as @p input.
+     * @param[in] output_state_out            2D weights tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
+     * @param[in] cell_state_out              2D tensor info with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
+     * @param[in] output                      Destination tensor info. Output is a 2D tensor with dimensions [output_size, batch_size]. Data types supported: Same as @p input.
+     * @param[in] lstm_params                 (Optional) Weights tensors info used in peephole optimization:
+     *                                        input_to_input_weights         2D weights tensor info with dimensions [input_size, num_units]. Data type supported: Same as @p input.
+     *                                        recurrent_to_input_weights     2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     *                                        cell_to_input_weights          1D weights tensor info with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
+     *                                        cell_to_forget_weights         1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        cell_to_output_weights         1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        input_gate_bias                1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input
+     *                                        projection_weights             2D weights tensor info with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     *                                        projection_bias                1D weights tensor info with dimensions [output_size]. Data type supported: Same as @p input.
+     *                                        input_layer_norm_coefficients  1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        forget_layer_norm_coefficients 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        cell_layer_norm_coefficients   1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        output_layer_norm_coefficients 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
      * @param[in] activation_info             Contains activation information described in @ref ActivationLayerInfo.
-     * @param[in] cell_threshold              The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. If set to 0.0 then clipping is disabled.
-     * @param[in] projection_threshold        The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled.
+     * @param[in] cell_threshold              The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip]. If set to 0.0f then clipping is disabled.
+     * @param[in] projection_threshold        The clipping threshold for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip].
+     *                                        If set to 0.0f then clipping is disabled.
      *
      * @return a status
      */
@@ -145,23 +156,16 @@ public:
 private:
     CLMemoryGroup                        _memory_group;
     CLFullyConnectedLayer                _fully_connected_input_gate;
-    CLGEMM                               _gemm_input_gate;
-    CLTransposeKernel                    _transpose_input_gate;
-    CLSaturatedArithmeticOperationKernel _accum_input_gate1;
-    CLArithmeticAddition                 _accum_input_gate2;
+    CLArithmeticAddition                 _accum_input_gate1;
     CLSaturatedArithmeticOperationKernel _subtract_input_gate;
     CLPixelWiseMultiplicationKernel      _pixelwise_mul_input_gate;
     CLActivationLayerKernel              _activation_input_gate;
     CLFullyConnectedLayer                _fully_connected_forget_gate;
-    CLGEMM                               _gemm_forget_gate;
-    CLTransposeKernel                    _transpose_forget_gate;
-    CLSaturatedArithmeticOperationKernel _accum_forget_gate1;
-    CLArithmeticAddition                 _accum_forget_gate2;
+    CLArithmeticAddition                 _accum_forget_gate1;
     CLPixelWiseMultiplicationKernel      _pixelwise_mul_forget_gate;
     CLActivationLayerKernel              _activation_forget_gate;
     CLFullyConnectedLayer                _fully_connected_cell_state;
     CLGEMM                               _gemm_cell_state1;
-    CLGEMM                               _gemm_cell_state2;
     CLTransposeKernel                    _transpose_cell_state;
     CLSaturatedArithmeticOperationKernel _accum_cell_state1;
     CLSaturatedArithmeticOperationKernel _accum_cell_state2;
@@ -170,17 +174,12 @@ private:
     CLActivationLayerKernel              _cell_clip;
     CLPixelWiseMultiplicationKernel      _pixelwise_mul_cell_state2;
     CLFullyConnectedLayer                _fully_connected_output;
-    CLGEMM                               _gemm_output;
     CLPixelWiseMultiplicationKernel      _pixelwise_mul_output_state1;
-    CLTransposeKernel                    _transpose_output;
-    CLSaturatedArithmeticOperationKernel _accum_output1;
-    CLArithmeticAddition                 _accum_output2;
+    CLArithmeticAddition                 _accum_output1;
     CLActivationLayerKernel              _activation_output;
     CLActivationLayerKernel              _activation_output_state;
     CLPixelWiseMultiplicationKernel      _pixelwise_mul_output_state2;
     CLFullyConnectedLayer                _fully_connected_output_state;
-    CLGEMM                               _gemm_output_state;
-    CLSaturatedArithmeticOperationKernel _accum_output_state;
     CLActivationLayerKernel              _projection_clip;
     CLCopyKernel                         _copy_cell_state;
     CLCopyKernel                         _copy_output;
@@ -190,6 +189,18 @@ private:
     CLWidthConcatenate2TensorsKernel     _concat_weights_input_gate;
     CLWidthConcatenate2TensorsKernel     _concat_weights_output;
     CLMemsetKernel                       _ones_memset_kernel;
+    CLMeanStdDevNormalizationLayer       _mean_std_norm_input_gate;
+    CLPixelWiseMultiplicationKernel      _pixelwise_mul_input_gate_coeff;
+    CLSaturatedArithmeticOperationKernel _accum_input_gate_bias;
+    CLMeanStdDevNormalizationLayer       _mean_std_norm_forget_gate;
+    CLPixelWiseMultiplicationKernel      _pixelwise_mul_forget_gate_coeff;
+    CLSaturatedArithmeticOperationKernel _accum_forget_gate_bias;
+    CLMeanStdDevNormalizationLayer       _mean_std_norm_cell_gate;
+    CLPixelWiseMultiplicationKernel      _pixelwise_mul_cell_gate_coeff;
+    CLSaturatedArithmeticOperationKernel _accum_cell_gate_bias;
+    CLMeanStdDevNormalizationLayer       _mean_std_norm_output_gate;
+    CLPixelWiseMultiplicationKernel      _pixelwise_mul_output_gate_coeff;
+    CLSaturatedArithmeticOperationKernel _accum_output_gate_bias;
     CLTensor                             _input_gate_out1;
     CLTensor                             _input_gate_out2;
     CLTensor                             _input_gate_out3;
@@ -212,12 +223,21 @@ private:
     CLTensor                             _cell_state_activation;
     CLTensor                             _output_state1;
     CLTensor                             _ones;
+    CLTensor                             _input_layer_norm_out1;
+    CLTensor                             _input_layer_norm_out2;
+    CLTensor                             _forget_layer_norm_out1;
+    CLTensor                             _forget_layer_norm_out2;
+    CLTensor                             _cell_layer_norm_out1;
+    CLTensor                             _cell_layer_norm_out2;
+    CLTensor                             _output_layer_norm_out1;
+    CLTensor                             _output_layer_norm_out2;
     bool                                 _run_peephole_opt;
     bool                                 _run_cifg_opt;
     bool                                 _perform_cell_clipping;
     bool                                 _has_projection_weights;
     bool                                 _perform_projection_clipping;
     bool                                 _is_prepared;
+    bool                                 _is_layer_norm_lstm;
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_CLLSTMLAYER_H__ */
diff --git a/arm_compute/runtime/common/LSTMParams.h b/arm_compute/runtime/common/LSTMParams.h
index 5b33e2e937..6979f90721 100644
--- a/arm_compute/runtime/common/LSTMParams.h
+++ b/arm_compute/runtime/common/LSTMParams.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,7 +41,8 @@ public:
     /** Constructor */
     LSTMParams()
         : _input_to_input_weights(nullptr), _recurrent_to_input_weights(nullptr), _cell_to_input_weights(nullptr), _input_gate_bias(nullptr), _cell_to_forget_weights(nullptr),
-          _cell_to_output_weights(nullptr), _projection_weights(nullptr), _projection_bias(nullptr), _has_peephole_opt(false), _has_projection(false), _has_cifg_opt(true)
+          _cell_to_output_weights(nullptr), _projection_weights(nullptr), _projection_bias(nullptr), _input_layer_norm_weights(nullptr), _forget_layer_norm_weights(nullptr), _cell_layer_norm_weights(nullptr),
+          _output_layer_norm_weights(nullptr), _has_peephole_opt(false), _has_projection(false), _has_cifg_opt(true), _use_layer_norm(false)
     {
     }
     /** Prevent instances of this class from being copied (As this class contains pointers) */
@@ -96,6 +97,25 @@ public:
         _has_peephole_opt       = true;
         return *this;
     }
+    /** Set layer normalization tensor parameters.
+     *
+     * @param[in] input_layer_norm_weights  1D weights tensor with dimensions [num_units]. Data type supported: Data types supported: F16/F32.
+     * @param[in] forget_layer_norm_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input_layer_norm_weights.
+     * @param[in] cell_layer_norm_weights   1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input_layer_norm_weights.
+     * @param[in] output_layer_norm_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input_layer_norm_weights.
+     *
+     * @return Reference to this LSTMParams object
+     */
+    LSTMParams &set_layer_normalization_params(const T *input_layer_norm_weights, const T *forget_layer_norm_weights,
+                                               const T *cell_layer_norm_weights, const T *output_layer_norm_weights)
+    {
+        _input_layer_norm_weights  = input_layer_norm_weights;
+        _forget_layer_norm_weights = forget_layer_norm_weights;
+        _cell_layer_norm_weights   = cell_layer_norm_weights;
+        _output_layer_norm_weights = output_layer_norm_weights;
+        _use_layer_norm            = true;
+        return *this;
+    }
 
     const T *input_to_input_weights() const
     {
@@ -137,6 +157,26 @@ public:
         return _projection_bias;
     }
 
+    const T *input_layer_norm_weights() const
+    {
+        return _input_layer_norm_weights;
+    }
+
+    const T *forget_layer_norm_weights() const
+    {
+        return _forget_layer_norm_weights;
+    }
+
+    const T *cell_layer_norm_weights() const
+    {
+        return _cell_layer_norm_weights;
+    }
+
+    const T *output_layer_norm_weights() const
+    {
+        return _output_layer_norm_weights;
+    }
+
     bool has_peephole_opt() const
     {
         return _has_peephole_opt;
@@ -152,6 +192,11 @@ public:
         return _has_cifg_opt;
     }
 
+    bool use_layer_norm() const
+    {
+        return _use_layer_norm;
+    }
+
 private:
     const T *_input_to_input_weights;
     const T *_recurrent_to_input_weights;
@@ -161,9 +206,14 @@ private:
     const T *_cell_to_output_weights;
     const T *_projection_weights;
     const T *_projection_bias;
+    const T *_input_layer_norm_weights;
+    const T *_forget_layer_norm_weights;
+    const T *_cell_layer_norm_weights;
+    const T *_output_layer_norm_weights;
     bool     _has_peephole_opt;
     bool     _has_projection;
     bool     _has_cifg_opt;
+    bool     _use_layer_norm;
 };
 }
 #endif /*__ARM_COMPUTE_LSTMPARAMS_H__ */
diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp
index 85a81a8cd4..793d5ca1a9 100644
--- a/src/runtime/CL/functions/CLLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayer.cpp
@@ -38,16 +38,18 @@ using namespace arm_compute;
 using namespace arm_compute::misc::shape_calculator;
 
 CLLSTMLayer::CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _gemm_input_gate(), _transpose_input_gate(), _accum_input_gate1(), _accum_input_gate2(), _subtract_input_gate(),
-      _pixelwise_mul_input_gate(), _activation_input_gate(), _fully_connected_forget_gate(), _gemm_forget_gate(), _transpose_forget_gate(), _accum_forget_gate1(), _accum_forget_gate2(),
-      _pixelwise_mul_forget_gate(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(), _gemm_cell_state2(), _transpose_cell_state(), _accum_cell_state1(), _accum_cell_state2(),
-      _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(), _pixelwise_mul_cell_state2(), _fully_connected_output(), _gemm_output(), _pixelwise_mul_output_state1(), _transpose_output(),
-      _accum_output1(), _accum_output2(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(), _fully_connected_output_state(), _gemm_output_state(), _accum_output_state(),
-      _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _concat_inputs_forget_gate(), _concat_weights_forget_gate(), _concat_weights_input_gate(), _concat_weights_output(),
-      _ones_memset_kernel(), _input_gate_out1(), _input_gate_out2(), _input_gate_out3(), _input_gate_out4(), _forget_gate_out1(), _forget_gate_out2(), _forget_gate_out3(), _forget_gate_out4(),
+    : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _accum_input_gate1(), _subtract_input_gate(), _pixelwise_mul_input_gate(), _activation_input_gate(),
+      _fully_connected_forget_gate(), _accum_forget_gate1(), _pixelwise_mul_forget_gate(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(), _transpose_cell_state(),
+      _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(), _pixelwise_mul_cell_state2(), _fully_connected_output(),
+      _pixelwise_mul_output_state1(), _accum_output1(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(), _fully_connected_output_state(), _projection_clip(),
+      _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _concat_inputs_forget_gate(), _concat_weights_forget_gate(), _concat_weights_input_gate(), _concat_weights_output(),
+      _ones_memset_kernel(), _mean_std_norm_input_gate(), _pixelwise_mul_input_gate_coeff(), _accum_input_gate_bias(), _mean_std_norm_forget_gate(), _pixelwise_mul_forget_gate_coeff(),
+      _accum_forget_gate_bias(), _mean_std_norm_cell_gate(), _pixelwise_mul_cell_gate_coeff(), _accum_cell_gate_bias(), _mean_std_norm_output_gate(), _pixelwise_mul_output_gate_coeff(),
+      _accum_output_gate_bias(), _input_gate_out1(), _input_gate_out2(), _input_gate_out3(), _input_gate_out4(), _forget_gate_out1(), _forget_gate_out2(), _forget_gate_out3(), _forget_gate_out4(),
       _forget_gate_out5(), _forget_gate_out6(), _cell_state_out1(), _cell_state_out2(), _cell_state_out3(), _cell_state_out4(), _cell_state_out5(), _output1(), _output2(), _output3(), _output4(),
-      _cell_state_activation(), _output_state1(), _ones(), _run_peephole_opt(false), _run_cifg_opt(false), _perform_cell_clipping(false), _has_projection_weights(false), _perform_projection_clipping(false),
-      _is_prepared(false)
+      _cell_state_activation(), _output_state1(), _ones(), _input_layer_norm_out1(), _input_layer_norm_out2(), _forget_layer_norm_out1(), _forget_layer_norm_out2(), _cell_layer_norm_out1(),
+      _cell_layer_norm_out2(), _output_layer_norm_out1(), _output_layer_norm_out2(), _run_peephole_opt(false), _run_cifg_opt(false), _perform_cell_clipping(false), _has_projection_weights(false),
+      _perform_projection_clipping(false), _is_prepared(false), _is_layer_norm_lstm(false)
 {
 }
 
@@ -66,6 +68,8 @@ void CLLSTMLayer::configure(const ICLTensor *input,
                                  output_state_in, cell_state_in,
                                  scratch_buffer, output_state_out, cell_state_out, output);
 
+    _is_layer_norm_lstm = lstm_params.use_layer_norm();
+
     // Set lstm parameters
     LSTMParams<ITensorInfo> lstm_params_info;
     if(lstm_params.has_peephole_opt())
@@ -121,7 +125,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
     _concat_weights_forget_gate.configure(input_to_forget_weights, recurrent_to_forget_weights, &_forget_gate_out6);
 
     _memory_group.manage(&_forget_gate_out5);
-    _fully_connected_forget_gate.configure(&_forget_gate_out2, &_forget_gate_out6, forget_gate_bias, &_forget_gate_out5);
+    _fully_connected_forget_gate.configure(&_forget_gate_out2, &_forget_gate_out6, (_is_layer_norm_lstm) ? nullptr : forget_gate_bias, &_forget_gate_out5);
     _memory_group.manage(&_forget_gate_out1);
     _memory_group.manage(&_forget_gate_out3);
     _forget_gate_out6.allocator()->allocate();
@@ -134,7 +138,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
         _run_peephole_opt = true;
         _memory_group.manage(&_forget_gate_out4);
         _pixelwise_mul_forget_gate.configure(cell_state_in, lstm_params.cell_to_forget_weights(), &_forget_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-        _accum_forget_gate2.configure(&_forget_gate_out5, &_forget_gate_out4, &_forget_gate_out3, ConvertPolicy::SATURATE);
+        _accum_forget_gate1.configure(&_forget_gate_out5, &_forget_gate_out4, &_forget_gate_out3, ConvertPolicy::SATURATE);
         _forget_gate_out4.allocator()->allocate();
         _forget_gate_out5.allocator()->allocate();
         forget_gate_out = &_forget_gate_out3;
@@ -143,6 +147,20 @@ void CLLSTMLayer::configure(const ICLTensor *input,
     {
         _forget_gate_out3.allocator()->allocate();
     }
+    if(_is_layer_norm_lstm)
+    {
+        _forget_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+        _forget_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+        _memory_group.manage(&_forget_layer_norm_out1);
+        _memory_group.manage(&_forget_layer_norm_out2);
+        _mean_std_norm_forget_gate.configure(forget_gate_out);
+        _pixelwise_mul_forget_gate_coeff.configure(forget_gate_out, lstm_params.forget_layer_norm_weights(), &_forget_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+        // forget_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
+        forget_gate_out->allocator()->allocate();
+        _accum_forget_gate_bias.configure(ArithmeticOperation::ADD, &_forget_layer_norm_out1, forget_gate_bias, &_forget_layer_norm_out2, ConvertPolicy::SATURATE);
+        _forget_layer_norm_out1.allocator()->allocate();
+        forget_gate_out = &_forget_layer_norm_out2;
+    }
     _activation_forget_gate.configure(forget_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
 
     // Configure block that calculates the input gate
@@ -177,7 +195,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
         _memory_group.manage(&_input_gate_out1);
 
         _memory_group.manage(&_input_gate_out3);
-        _fully_connected_input_gate.configure(&_forget_gate_out2, &_input_gate_out2, lstm_params.input_gate_bias(), &_input_gate_out3);
+        _fully_connected_input_gate.configure(&_forget_gate_out2, &_input_gate_out2, (_is_layer_norm_lstm) ? nullptr : lstm_params.input_gate_bias(), &_input_gate_out3);
         _input_gate_out2.allocator()->allocate();
 
         input_gate_out = &_input_gate_out3;
@@ -185,7 +203,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
         {
             _memory_group.manage(&_input_gate_out4);
             _pixelwise_mul_input_gate.configure(cell_state_in, lstm_params.cell_to_input_weights(), &_input_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-            _accum_input_gate2.configure(&_input_gate_out3, &_input_gate_out4, &_input_gate_out1, ConvertPolicy::SATURATE);
+            _accum_input_gate1.configure(&_input_gate_out3, &_input_gate_out4, &_input_gate_out1, ConvertPolicy::SATURATE);
             _input_gate_out3.allocator()->allocate();
             _input_gate_out4.allocator()->allocate();
             input_gate_out = &_input_gate_out1;
@@ -194,6 +212,21 @@ void CLLSTMLayer::configure(const ICLTensor *input,
         {
             _input_gate_out1.allocator()->allocate();
         }
+
+        if(_is_layer_norm_lstm)
+        {
+            _input_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+            _input_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+            _memory_group.manage(&_input_layer_norm_out1);
+            _memory_group.manage(&_input_layer_norm_out2);
+            _mean_std_norm_input_gate.configure(input_gate_out);
+            _pixelwise_mul_input_gate_coeff.configure(input_gate_out, lstm_params.input_layer_norm_weights(), &_input_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+            // input_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
+            input_gate_out->allocator()->allocate();
+            _accum_input_gate_bias.configure(ArithmeticOperation::ADD, &_input_layer_norm_out1, lstm_params.input_gate_bias(), &_input_layer_norm_out2, ConvertPolicy::SATURATE);
+            _input_layer_norm_out1.allocator()->allocate();
+            input_gate_out = &_input_layer_norm_out2;
+        }
         _activation_input_gate.configure(input_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
     }
 
@@ -207,7 +240,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
     _cell_state_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
 
     _memory_group.manage(&_cell_state_out1);
-    _fully_connected_cell_state.configure(input, input_to_cell_weights, cell_bias, &_cell_state_out1);
+    _fully_connected_cell_state.configure(input, input_to_cell_weights, (_is_layer_norm_lstm) ? nullptr : cell_bias, &_cell_state_out1);
     _memory_group.manage(&_cell_state_out2);
     _transpose_cell_state.configure(recurrent_to_cell_weights, &_cell_state_out2);
     _memory_group.manage(&_cell_state_out3);
@@ -215,10 +248,25 @@ void CLLSTMLayer::configure(const ICLTensor *input,
     _cell_state_out2.allocator()->allocate();
     _memory_group.manage(&_cell_state_out4);
     _accum_cell_state1.configure(ArithmeticOperation::ADD, &_cell_state_out1, &_cell_state_out3, &_cell_state_out4, ConvertPolicy::SATURATE);
-    _activation_cell_state.configure(&_cell_state_out4, nullptr, activation_info);
+    CLTensor *cell_state_out_ptr = &_cell_state_out4;
+    if(_is_layer_norm_lstm)
+    {
+        _cell_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+        _cell_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+        _memory_group.manage(&_cell_layer_norm_out1);
+        _memory_group.manage(&_cell_layer_norm_out2);
+        _mean_std_norm_cell_gate.configure(cell_state_out_ptr);
+        _pixelwise_mul_cell_gate_coeff.configure(cell_state_out_ptr, lstm_params.cell_layer_norm_weights(), &_cell_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+        // cell_state_out_ptr is going to be reassigned, so allocate the tensor that it was assigned to before
+        cell_state_out_ptr->allocator()->allocate();
+        _accum_cell_gate_bias.configure(ArithmeticOperation::ADD, &_cell_layer_norm_out1, cell_bias, &_cell_layer_norm_out2, ConvertPolicy::SATURATE);
+        _cell_layer_norm_out1.allocator()->allocate();
+        cell_state_out_ptr = &_cell_layer_norm_out2;
+    }
+    _activation_cell_state.configure(cell_state_out_ptr, nullptr, activation_info);
     _memory_group.manage(&_cell_state_out5);
-    _pixelwise_mul_cell_state1.configure(&_cell_state_out4, input_gate_out, &_cell_state_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-    _cell_state_out4.allocator()->allocate();
+    _pixelwise_mul_cell_state1.configure(cell_state_out_ptr, input_gate_out, &_cell_state_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+    cell_state_out_ptr->allocator()->allocate();
     _pixelwise_mul_cell_state2.configure(forget_gate_out, cell_state_in, &_cell_state_out3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
     _accum_cell_state2.configure(ArithmeticOperation::ADD, &_cell_state_out5, &_cell_state_out3, &_cell_state_out1, ConvertPolicy::SATURATE);
     _cell_state_out3.allocator()->allocate();
@@ -247,7 +295,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
     _memory_group.manage(&_output1);
     _memory_group.manage(&_output4);
 
-    _fully_connected_output.configure(&_forget_gate_out2, &_output2, output_gate_bias, &_output4);
+    _fully_connected_output.configure(&_forget_gate_out2, &_output2, (_is_layer_norm_lstm) ? nullptr : output_gate_bias, &_output4);
 
     _output2.allocator()->allocate();
     _forget_gate_out2.allocator()->allocate();
@@ -259,7 +307,7 @@ void CLLSTMLayer::configure(const ICLTensor *input,
 
         _memory_group.manage(&_output3);
         _pixelwise_mul_output_state1.configure(&_cell_state_out1, lstm_params.cell_to_output_weights(), &_output3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-        _accum_output2.configure(&_output4, &_output3, &_output1, ConvertPolicy::SATURATE);
+        _accum_output1.configure(&_output4, &_output3, &_output1, ConvertPolicy::SATURATE);
         _output4.allocator()->allocate();
         output_gate_out = &_output1;
 
@@ -270,6 +318,20 @@ void CLLSTMLayer::configure(const ICLTensor *input,
     {
         _output1.allocator()->allocate();
     }
+    if(_is_layer_norm_lstm)
+    {
+        _output_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+        _output_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+        _memory_group.manage(&_output_layer_norm_out1);
+        _memory_group.manage(&_output_layer_norm_out2);
+        _mean_std_norm_output_gate.configure(output_gate_out);
+        _pixelwise_mul_output_gate_coeff.configure(output_gate_out, lstm_params.output_layer_norm_weights(), &_output_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+        // output_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
+        output_gate_out->allocator()->allocate();
+        _accum_output_gate_bias.configure(ArithmeticOperation::ADD, &_output_layer_norm_out1, output_gate_bias, &_output_layer_norm_out2, ConvertPolicy::SATURATE);
+        _output_layer_norm_out1.allocator()->allocate();
+        output_gate_out = &_output_layer_norm_out2;
+    }
     _activation_output.configure(output_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
 
     // Configure block that calculates the output state
@@ -370,6 +432,31 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
     const unsigned int num_batches = input->dimension(1);
     const unsigned int num_cells   = input_to_output_weights->dimension(1);
 
+    if(lstm_params.use_layer_norm())
+    {
+        // If CIFG is used, input layer normalization weights tensor is omitted
+        if(lstm_params.has_cifg_opt())
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights() != nullptr);
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.input_layer_norm_weights());
+            ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->num_dimensions() > 1);
+            ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->dimension(0) != num_batches);
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, lstm_params.input_layer_norm_weights());
+        }
+
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.forget_layer_norm_weights(), lstm_params.cell_layer_norm_weights(), lstm_params.output_layer_norm_weights());
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, lstm_params.forget_layer_norm_weights(), lstm_params.cell_layer_norm_weights(), lstm_params.output_layer_norm_weights());
+        ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->dimension(0) != num_batches);
+        ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->dimension(0) != num_batches);
+        ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->dimension(0) != num_batches);
+    }
+
     // Check peephole optimization
     if(lstm_params.has_peephole_opt())
     {
@@ -389,7 +476,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
     TensorInfo cell_state_tmp  = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
 
     // Validate forget gate
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, forget_gate_bias, &forget_gate));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, (lstm_params.use_layer_norm()) ? nullptr : forget_gate_bias, &forget_gate));
 
     std::vector<const ITensorInfo *> inputs_vector;
     inputs_vector.emplace_back(input);
@@ -404,6 +491,13 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
         ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE));
     }
+    if(lstm_params.use_layer_norm())
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&forget_gate));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&forget_gate, lstm_params.forget_layer_norm_weights(), &forget_gate, 1, ConvertPolicy::SATURATE,
+                                                                              RoundingPolicy::TO_NEAREST_EVEN));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE));
+    }
     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
 
     // Validate input gate
@@ -423,7 +517,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
         TensorInfo  lstm_gate_concat          = TensorInfo(lstm_weights_concat_shape, 1, input->data_type());
         ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenate2TensorsKernel::validate(lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights(), &lstm_gate_concat));
 
-        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), lstm_params.input_gate_bias(), &input_gate));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), (lstm_params.use_layer_norm()) ? nullptr : lstm_params.input_gate_bias(), &input_gate));
 
         if(lstm_params.has_peephole_opt())
         {
@@ -432,6 +526,13 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
             ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
             ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE));
         }
+
+        if(lstm_params.use_layer_norm())
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&input_gate));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&input_gate, lstm_params.input_layer_norm_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, lstm_params.input_gate_bias(), &input_gate, ConvertPolicy::SATURATE));
+        }
         ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
     }
     else
@@ -440,9 +541,16 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
     }
 
     // Validate cell state
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, cell_bias, &cell_state_tmp));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, (lstm_params.use_layer_norm()) ? nullptr : cell_bias, &cell_state_tmp));
     ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &cell_state_tmp, 1.f, 0.f, GEMMInfo()));
     ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
+    if(lstm_params.use_layer_norm())
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&cell_state_tmp));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, lstm_params.cell_layer_norm_weights(), &cell_state_tmp, 1, ConvertPolicy::SATURATE,
+                                                                              RoundingPolicy::TO_NEAREST_EVEN));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE));
+    }
     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&cell_state_tmp, nullptr, activation_info));
     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
@@ -460,7 +568,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
     TensorInfo  in_out_gate_concat          = TensorInfo(in_out_weights_concat_shape, 1, input->data_type());
     ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenate2TensorsKernel::validate(input_to_output_weights, recurrent_to_output_weights, &in_out_gate_concat));
     // Validate output gate tmp
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, output_gate_bias, &output_gate_tmp));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, (lstm_params.use_layer_norm()) ? nullptr : output_gate_bias, &output_gate_tmp));
 
     if(lstm_params.has_peephole_opt())
     {
@@ -468,6 +576,13 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
                                                                               RoundingPolicy::TO_NEAREST_EVEN));
         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE));
     }
+    if(lstm_params.use_layer_norm())
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&output_gate_tmp));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&output_gate_tmp, lstm_params.output_layer_norm_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE,
+                                                                              RoundingPolicy::TO_NEAREST_EVEN));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, ConvertPolicy::SATURATE));
+    }
     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
 
     // Validate output state
@@ -514,7 +629,13 @@ void CLLSTMLayer::run()
     if(_run_peephole_opt)
     {
         CLScheduler::get().enqueue(_pixelwise_mul_forget_gate);
-        _accum_forget_gate2.run();
+        _accum_forget_gate1.run();
+    }
+    if(_is_layer_norm_lstm)
+    {
+        _mean_std_norm_forget_gate.run();
+        CLScheduler::get().enqueue(_pixelwise_mul_forget_gate_coeff);
+        CLScheduler::get().enqueue(_accum_forget_gate_bias);
     }
     CLScheduler::get().enqueue(_activation_forget_gate);
 
@@ -530,7 +651,14 @@ void CLLSTMLayer::run()
         if(_run_peephole_opt)
         {
             CLScheduler::get().enqueue(_pixelwise_mul_input_gate);
-            _accum_input_gate2.run();
+            _accum_input_gate1.run();
+        }
+
+        if(_is_layer_norm_lstm)
+        {
+            _mean_std_norm_input_gate.run();
+            CLScheduler::get().enqueue(_pixelwise_mul_input_gate_coeff);
+            CLScheduler::get().enqueue(_accum_input_gate_bias);
         }
         CLScheduler::get().enqueue(_activation_input_gate);
     }
@@ -539,6 +667,12 @@ void CLLSTMLayer::run()
     CLScheduler::get().enqueue(_transpose_cell_state);
     _gemm_cell_state1.run();
     CLScheduler::get().enqueue(_accum_cell_state1);
+    if(_is_layer_norm_lstm)
+    {
+        _mean_std_norm_cell_gate.run();
+        CLScheduler::get().enqueue(_pixelwise_mul_cell_gate_coeff);
+        CLScheduler::get().enqueue(_accum_cell_gate_bias);
+    }
     CLScheduler::get().enqueue(_activation_cell_state);
     CLScheduler::get().enqueue(_pixelwise_mul_cell_state1);
     CLScheduler::get().enqueue(_pixelwise_mul_cell_state2);
@@ -554,7 +688,13 @@ void CLLSTMLayer::run()
     if(_run_peephole_opt)
     {
         CLScheduler::get().enqueue(_pixelwise_mul_output_state1);
-        _accum_output2.run();
+        _accum_output1.run();
+    }
+    if(_is_layer_norm_lstm)
+    {
+        _mean_std_norm_output_gate.run();
+        CLScheduler::get().enqueue(_pixelwise_mul_output_gate_coeff);
+        CLScheduler::get().enqueue(_accum_output_gate_bias);
     }
     CLScheduler::get().enqueue(_activation_output);
 
diff --git a/tests/validation/CL/LSTMLayer.cpp b/tests/validation/CL/LSTMLayer.cpp
index 71a9383d93..69ac61dcf4 100644
--- a/tests/validation/CL/LSTMLayer.cpp
+++ b/tests/validation/CL/LSTMLayer.cpp
@@ -153,10 +153,11 @@ template <typename T>
 using CLLSTMLayerFixture = LSTMLayerValidationFixture<CLTensor, CLAccessor, CLLSTMLayer, LSTMParams<ICLTensor>, T>;
 
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLLSTMLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallLSTMLayerDataset(), framework::dataset::make("DataType",
+FIXTURE_DATA_TEST_CASE(RunSmall, CLLSTMLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallLSTMLayerDataset(), framework::dataset::make("DataType",
                                                                                                                  DataType::F32)),
-                                                                                                         framework::dataset::make("ProjectionOpt", { true, false })),
-                                                                                                 framework::dataset::make("PeepholeOpt", { true, false })))
+                                                                                                                 framework::dataset::make("ProjectionOpt", { true, false })),
+                                                                                                         framework::dataset::make("PeepholeOpt", { true, false })),
+                                                                                                 framework::dataset::make("UseLayerNorm", { true, false })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32);
@@ -165,9 +166,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLLSTMLayerFixture<float>, framework::DatasetMo
 TEST_SUITE_END() // FP32
 
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLLSTMLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallLSTMLayerDataset(), framework::dataset::make("DataType", DataType::F16)),
-                                                                                                        framework::dataset::make("ProjectionOpt", { true, false })),
-                                                                                                framework::dataset::make("PeepholeOpt", { true, false })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLLSTMLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallLSTMLayerDataset(), framework::dataset::make("DataType",
+                                                                                                                        DataType::F16)),
+                                                                                                                framework::dataset::make("ProjectionOpt", { true, false })),
+                                                                                                        framework::dataset::make("PeepholeOpt", { true, false })),
+                                                                                                framework::dataset::make("UseLayerNorm", { true, false })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f16);
diff --git a/tests/validation/NEON/LSTMLayer.cpp b/tests/validation/NEON/LSTMLayer.cpp
index b27dfae8fa..c503972ba9 100644
--- a/tests/validation/NEON/LSTMLayer.cpp
+++ b/tests/validation/NEON/LSTMLayer.cpp
@@ -153,10 +153,11 @@ template <typename T>
 using NELSTMLayerFixture = LSTMLayerValidationFixture<Tensor, Accessor, NELSTMLayer, LSTMParams<ITensor>, T>;
 
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, NELSTMLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallLSTMLayerDataset(), framework::dataset::make("DataType",
+FIXTURE_DATA_TEST_CASE(RunSmall, NELSTMLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallLSTMLayerDataset(), framework::dataset::make("DataType",
                                                                                                                  DataType::F32)),
-                                                                                                         framework::dataset::make("ProjectionOpt", { true, false })),
-                                                                                                 framework::dataset::make("PeepholeOpt", { true, false })))
+                                                                                                                 framework::dataset::make("ProjectionOpt", { true, false })),
+                                                                                                         framework::dataset::make("PeepholeOpt", { true, false })),
+                                                                                                 framework::dataset::make("UseLayerNorm", { false })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f32);
@@ -166,9 +167,11 @@ TEST_SUITE_END() // FP32
 
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NELSTMLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallLSTMLayerDataset(), framework::dataset::make("DataType", DataType::F16)),
-                                                                                                        framework::dataset::make("ProjectionOpt", { true, false })),
-                                                                                                framework::dataset::make("PeepholeOpt", { true, false })))
+FIXTURE_DATA_TEST_CASE(RunSmall, NELSTMLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallLSTMLayerDataset(), framework::dataset::make("DataType",
+                                                                                                                        DataType::F16)),
+                                                                                                                framework::dataset::make("ProjectionOpt", { true, false })),
+                                                                                                        framework::dataset::make("PeepholeOpt", { true, false })),
+                                                                                                framework::dataset::make("UseLayerNorm", { false })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f16);
diff --git a/tests/validation/fixtures/LSTMLayerFixture.h b/tests/validation/fixtures/LSTMLayerFixture.h
index 2cf83b8b3d..9260686d56 100644
--- a/tests/validation/fixtures/LSTMLayerFixture.h
+++ b/tests/validation/fixtures/LSTMLayerFixture.h
@@ -32,6 +32,7 @@
 #include "tests/validation/reference/ConcatenateLayer.h"
 #include "tests/validation/reference/FullyConnectedLayer.h"
 #include "tests/validation/reference/GEMM.h"
+#include "tests/validation/reference/MeanStdDevNormalizationLayer.h"
 #include "tests/validation/reference/PixelWiseMultiplication.h"
 #include "tests/validation/reference/Transpose.h"
 
@@ -47,12 +48,13 @@ class LSTMLayerValidationFixture : public framework::Fixture
 public:
     template <typename...>
     void setup(TensorShape input_shape, TensorShape input_weights_shape, TensorShape recurrent_weights_shape, TensorShape cell_bias_shape, TensorShape output_cell_shape, TensorShape output_shape,
-               TensorShape scratch_shape, ActivationLayerInfo info, float cell_threshold, float projection_threshold, DataType data_type, bool projection_opt, bool peephole_opt)
+               TensorShape scratch_shape, ActivationLayerInfo info, float cell_threshold, float projection_threshold, DataType data_type, bool projection_opt, bool peephole_opt,
+               bool use_layer_norm)
     {
         _target = compute_target(input_shape, input_weights_shape, recurrent_weights_shape, cell_bias_shape, output_cell_shape, output_shape, scratch_shape, info, cell_threshold, projection_threshold,
-                                 data_type, projection_opt, peephole_opt);
+                                 data_type, projection_opt, peephole_opt, use_layer_norm);
         _reference = compute_reference(input_shape, input_weights_shape, recurrent_weights_shape, cell_bias_shape, output_cell_shape, output_shape, scratch_shape, info, cell_threshold, projection_threshold,
-                                       data_type, projection_opt, peephole_opt);
+                                       data_type, projection_opt, peephole_opt, use_layer_norm);
     }
 
 protected:
@@ -70,7 +72,7 @@ protected:
     }
     TensorType compute_target(const TensorShape &input_shape, const TensorShape &input_weights_shape, const TensorShape &recurrent_weights_shape, const TensorShape &cell_bias_shape,
                               const TensorShape &output_cell_shape, const TensorShape &output_shape, const TensorShape &scratch_shape, ActivationLayerInfo info, float cell_threshold,
-                              float projection_threshold, DataType data_type, bool projection_opt, bool peephole_opt)
+                              float projection_threshold, DataType data_type, bool projection_opt, bool peephole_opt, bool use_layer_norm)
     {
         const unsigned int num_cells   = input_weights_shape.y();
         const unsigned int num_outputs = recurrent_weights_shape.x();
@@ -100,6 +102,10 @@ protected:
         TensorType cell_to_output_w;
         TensorType projection_w;
         TensorType projection_bias;
+        TensorType input_layer_norm_w;
+        TensorType forget_layer_norm_w;
+        TensorType cell_layer_norm_w;
+        TensorType output_layer_norm_w;
 
         bool cifg_opt = scratch_shape.x() == cell_bias_shape.x() * 4 ? false : true;
 
@@ -131,6 +137,22 @@ protected:
             lstm_params.set_projection_params(&projection_w, &projection_bias);
         }
 
+        if(use_layer_norm)
+        {
+            forget_layer_norm_w = create_tensor<TensorType>(TensorShape(num_cells), data_type);
+            cell_layer_norm_w   = create_tensor<TensorType>(TensorShape(num_cells), data_type);
+            output_layer_norm_w = create_tensor<TensorType>(TensorShape(num_cells), data_type);
+            if(!cifg_opt)
+            {
+                input_layer_norm_w = create_tensor<TensorType>(TensorShape(num_cells), data_type);
+                lstm_params.set_layer_normalization_params(&input_layer_norm_w, &forget_layer_norm_w, &cell_layer_norm_w, &output_layer_norm_w);
+            }
+            else
+            {
+                lstm_params.set_layer_normalization_params(nullptr, &forget_layer_norm_w, &cell_layer_norm_w, &output_layer_norm_w);
+            }
+        }
+
         // Create and configure function
         FunctionType lstm;
         lstm.configure(&input, &input_to_forget_w, &input_to_cell_w, &input_to_output_w, &recurrent_to_forget_w,
@@ -257,6 +279,35 @@ protected:
             fill(AccessorType(projection_bias), 21);
         }
 
+        if(use_layer_norm)
+        {
+            if(!cifg_opt)
+            {
+                ARM_COMPUTE_EXPECT(input_layer_norm_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+                input_layer_norm_w.allocator()->allocate();
+
+                ARM_COMPUTE_EXPECT(!input_layer_norm_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+                fill(AccessorType(input_layer_norm_w), 22);
+            }
+            ARM_COMPUTE_EXPECT(forget_layer_norm_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+            ARM_COMPUTE_EXPECT(cell_layer_norm_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+            ARM_COMPUTE_EXPECT(output_layer_norm_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+            forget_layer_norm_w.allocator()->allocate();
+            cell_layer_norm_w.allocator()->allocate();
+            output_layer_norm_w.allocator()->allocate();
+
+            ARM_COMPUTE_EXPECT(!forget_layer_norm_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+            ARM_COMPUTE_EXPECT(!cell_layer_norm_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+            ARM_COMPUTE_EXPECT(!output_layer_norm_w.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+            fill(AccessorType(forget_layer_norm_w), 23);
+            fill(AccessorType(cell_layer_norm_w), 24);
+            fill(AccessorType(output_layer_norm_w), 25);
+        }
+
         // Compute function
         lstm.run();
 
@@ -266,7 +317,7 @@ protected:
 
     SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &input_weights_shape, const TensorShape &recurrent_weights_shape, const TensorShape &cell_bias_shape,
                                       const TensorShape &output_cell_shape, const TensorShape &output_shape, const TensorShape &scratch_shape, ActivationLayerInfo info, float cell_threshold,
-                                      float projection_threshold, DataType data_type, bool projection_opt, bool peephole_opt)
+                                      float projection_threshold, DataType data_type, bool projection_opt, bool peephole_opt, bool use_layer_norm)
     {
         const unsigned int num_cells   = input_weights_shape.y();
         const unsigned int num_outputs = recurrent_weights_shape.x();
@@ -306,6 +357,8 @@ protected:
         SimpleTensor<T> cell_state_out{ output_cell_shape, data_type };
         SimpleTensor<T> output{ output_shape, data_type };
 
+        bool cifg_opt = scratch_shape.x() == cell_bias_shape.x() * 4 ? false : true;
+
         // Fill reference
         fill(input, 0);
         fill(input_to_forget_w, 1);
@@ -314,9 +367,18 @@ protected:
         fill(recurrent_to_forget_w, 4);
         fill(recurrent_to_cell_w, 5);
         fill(recurrent_to_output_w, 6);
-        fill(forget_gate_bias, 7);
-        fill(cell_bias, 8);
-        fill(output_gate_bias, 9);
+        if(use_layer_norm)
+        {
+            fill_custom_val(forget_gate_bias, 0.f, 7);
+            fill_custom_val(cell_bias, 0.f, 8);
+            fill_custom_val(output_gate_bias, 0.f, 9);
+        }
+        else
+        {
+            fill(forget_gate_bias, 7);
+            fill(cell_bias, 8);
+            fill(output_gate_bias, 9);
+        }
         fill(output_state_in, 10);
         fill(cell_state_in, 11);
         fill(scratch, 12);
@@ -324,14 +386,19 @@ protected:
         fill(recurrent_to_input_w, 14);
         fill(cell_to_input_w, 15);
         fill(recurrent_to_input_w, 16);
-        fill(input_gate_bias, 17);
+        if(!cifg_opt && use_layer_norm)
+        {
+            fill_custom_val(input_gate_bias, 0.f, 17);
+        }
+        else
+        {
+            fill(input_gate_bias, 17);
+        }
         fill(cell_to_forget_w, 18);
         fill(cell_to_output_w, 19);
         fill(projection_w, 20);
         fill(projection_bias, 21);
 
-        bool cifg_opt = scratch_shape.x() == cell_bias_shape.x() * 4 ? false : true;
-
         // Compute forget_gate
         SimpleTensor<T> fully_connected_forget = reference::fully_connected_layer(input, input_to_forget_w, forget_gate_bias, output_cell_shape);
         SimpleTensor<T> transposed_weights     = reference::transpose(recurrent_to_forget_w);
@@ -344,6 +411,15 @@ protected:
             forget_gate                               = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, forget_gate, pixelwise_mul_forget_gate, data_type, ConvertPolicy::SATURATE);
         }
 
+        if(use_layer_norm)
+        {
+            SimpleTensor<T> forget_layer_norm_w{ cell_bias_shape, data_type };
+            fill(forget_layer_norm_w, 23);
+            forget_gate = reference::mean_std_normalization_layer(forget_gate);
+            forget_gate = reference::pixel_wise_multiplication(forget_gate, forget_layer_norm_w, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+            fill(forget_gate_bias, 7);
+            forget_gate = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, forget_gate, forget_gate_bias, data_type, ConvertPolicy::SATURATE);
+        }
         forget_gate = reference::activation_layer(forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
 
         // Compute input_gate
@@ -365,6 +441,15 @@ protected:
                 SimpleTensor<T> pixelwise_mul_input_gate = reference::pixel_wise_multiplication(cell_state_in, cell_to_input_w, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
                 input_gate                               = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, input_gate, pixelwise_mul_input_gate, data_type, ConvertPolicy::SATURATE);
             }
+            if(use_layer_norm)
+            {
+                SimpleTensor<T> input_layer_norm_w{ cell_bias_shape, data_type };
+                fill(input_layer_norm_w, 22);
+                input_gate = reference::mean_std_normalization_layer(input_gate);
+                input_gate = reference::pixel_wise_multiplication(input_gate, input_layer_norm_w, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+                fill(input_gate_bias, 17);
+                input_gate = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, input_gate, input_gate_bias, data_type, ConvertPolicy::SATURATE);
+            }
             input_gate = reference::activation_layer(input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
         }
 
@@ -374,9 +459,18 @@ protected:
         gemm                                       = reference::gemm(output_state_in, transposed_weights, cell_state_out, 1.f, 0.f);
         SimpleTensor<T> pixelwise_mul              = reference::pixel_wise_multiplication(cell_state_in, forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
         cell_state_out                             = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, fully_connected_cell_state, gemm, data_type, ConvertPolicy::SATURATE);
-        cell_state_out                             = reference::activation_layer(cell_state_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
-        cell_state_out                             = reference::pixel_wise_multiplication(cell_state_out, input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-        cell_state_out                             = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, cell_state_out, pixelwise_mul, data_type, ConvertPolicy::SATURATE);
+        if(use_layer_norm)
+        {
+            SimpleTensor<T> cell_layer_norm_w{ cell_bias_shape, data_type };
+            fill(cell_layer_norm_w, 24);
+            cell_state_out = reference::mean_std_normalization_layer(cell_state_out);
+            cell_state_out = reference::pixel_wise_multiplication(cell_state_out, cell_layer_norm_w, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+            fill(cell_bias, 8);
+            cell_state_out = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, cell_state_out, cell_bias, data_type, ConvertPolicy::SATURATE);
+        }
+        cell_state_out = reference::activation_layer(cell_state_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
+        cell_state_out = reference::pixel_wise_multiplication(cell_state_out, input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+        cell_state_out = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, cell_state_out, pixelwise_mul, data_type, ConvertPolicy::SATURATE);
         if(cell_threshold != 0.f)
         {
             cell_state_out = reference::activation_layer(cell_state_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, cell_threshold));
@@ -392,6 +486,15 @@ protected:
             pixelwise_mul = reference::pixel_wise_multiplication(cell_state_out, cell_to_output_w, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
             output        = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, output, pixelwise_mul, data_type, ConvertPolicy::SATURATE);
         }
+        if(use_layer_norm)
+        {
+            SimpleTensor<T> output_layer_norm_w{ cell_bias_shape, data_type };
+            fill(output_layer_norm_w, 25);
+            output = reference::mean_std_normalization_layer(output);
+            output = reference::pixel_wise_multiplication(output, output_layer_norm_w, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+            fill(output_gate_bias, 9);
+            output = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, output, output_gate_bias, data_type, ConvertPolicy::SATURATE);
+        }
         output = reference::activation_layer(output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
 
         // Compute output state