COMPMID-2343: Add layer normalization support in NELSTMLayer

Change-Id: I1f620d70c6eaadfb9e3a1b345de350ac0253b65c
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1366
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

3 files changed, 218 insertions, 59 deletions

diff --git a/arm_compute/runtime/NEON/functions/NELSTMLayer.h b/arm_compute/runtime/NEON/functions/NELSTMLayer.h
index cf0f06c215..183745c185 100644
--- a/arm_compute/runtime/NEON/functions/NELSTMLayer.h
+++ b/arm_compute/runtime/NEON/functions/NELSTMLayer.h
@@ -35,6 +35,7 @@
 #include "arm_compute/runtime/NEON/functions/NEConcatenateLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMM.h"
+#include "arm_compute/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.h"
 #include "arm_compute/runtime/common/LSTMParams.h"
 
 namespace arm_compute
@@ -68,14 +69,18 @@ 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.
@@ -108,14 +113,18 @@ public:
      * @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.
+     *                                        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 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.
@@ -137,23 +146,16 @@ public:
 private:
     MemoryGroup                     _memory_group;
     NEFullyConnectedLayer           _fully_connected_input_gate;
-    NEGEMM                          _gemm_input_gate;
-    NETransposeKernel               _transpose_input_gate;
-    NEArithmeticAdditionKernel      _accum_input_gate1;
-    NEArithmeticAddition            _accum_input_gate2;
+    NEArithmeticAddition            _accum_input_gate1;
     NEArithmeticSubtractionKernel   _subtract_input_gate;
     NEPixelWiseMultiplicationKernel _pixelwise_mul_input_gate;
     NEActivationLayerKernel         _activation_input_gate;
     NEFullyConnectedLayer           _fully_connected_forget_gate;
-    NEGEMM                          _gemm_forget_gate;
-    NETransposeKernel               _transpose_forget_gate;
-    NEArithmeticAdditionKernel      _accum_forget_gate1;
-    NEArithmeticAddition            _accum_forget_gate2;
+    NEArithmeticAddition            _accum_forget_gate1;
     NEPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate;
     NEActivationLayerKernel         _activation_forget_gate;
     NEFullyConnectedLayer           _fully_connected_cell_state;
     NEGEMM                          _gemm_cell_state1;
-    NEGEMM                          _gemm_cell_state2;
     NETransposeKernel               _transpose_cell_state;
     NEArithmeticAdditionKernel      _accum_cell_state1;
     NEArithmeticAdditionKernel      _accum_cell_state2;
@@ -162,17 +164,12 @@ private:
     NEActivationLayerKernel         _cell_clip;
     NEPixelWiseMultiplicationKernel _pixelwise_mul_cell_state2;
     NEFullyConnectedLayer           _fully_connected_output;
-    NEGEMM                          _gemm_output;
     NEPixelWiseMultiplicationKernel _pixelwise_mul_output_state1;
-    NETransposeKernel               _transpose_output;
-    NEArithmeticAdditionKernel      _accum_output1;
-    NEArithmeticAddition            _accum_output2;
+    NEArithmeticAddition            _accum_output1;
     NEActivationLayerKernel         _activation_output;
     NEActivationLayerKernel         _activation_output_state;
     NEPixelWiseMultiplicationKernel _pixelwise_mul_output_state2;
     NEFullyConnectedLayer           _fully_connected_output_state;
-    NEGEMM                          _gemm_output_state;
-    NEArithmeticAdditionKernel      _accum_output_state;
     NEActivationLayerKernel         _projection_clip;
     NECopyKernel                    _copy_cell_state;
     NECopyKernel                    _copy_output;
@@ -181,6 +178,18 @@ private:
     NEConcatenateLayer              _concat_weights_forget_gate;
     NEConcatenateLayer              _concat_weights_input_gate;
     NEConcatenateLayer              _concat_weights_output;
+    NEMeanStdDevNormalizationLayer  _mean_std_norm_input_gate;
+    NEPixelWiseMultiplicationKernel _pixelwise_mul_input_gate_coeff;
+    NEArithmeticAdditionKernel      _accum_input_gate_bias;
+    NEMeanStdDevNormalizationLayer  _mean_std_norm_forget_gate;
+    NEPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate_coeff;
+    NEArithmeticAdditionKernel      _accum_forget_gate_bias;
+    NEMeanStdDevNormalizationLayer  _mean_std_norm_cell_gate;
+    NEPixelWiseMultiplicationKernel _pixelwise_mul_cell_gate_coeff;
+    NEArithmeticAdditionKernel      _accum_cell_gate_bias;
+    NEMeanStdDevNormalizationLayer  _mean_std_norm_output_gate;
+    NEPixelWiseMultiplicationKernel _pixelwise_mul_output_gate_coeff;
+    NEArithmeticAdditionKernel      _accum_output_gate_bias;
     Tensor                          _input_gate_out1;
     Tensor                          _input_gate_out2;
     Tensor                          _input_gate_out3;
@@ -203,12 +212,21 @@ private:
     Tensor                          _cell_state_activation;
     Tensor                          _output_state1;
     Tensor                          _ones;
+    Tensor                          _input_layer_norm_out1;
+    Tensor                          _input_layer_norm_out2;
+    Tensor                          _forget_layer_norm_out1;
+    Tensor                          _forget_layer_norm_out2;
+    Tensor                          _cell_layer_norm_out1;
+    Tensor                          _cell_layer_norm_out2;
+    Tensor                          _output_layer_norm_out1;
+    Tensor                          _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_NELSTMLAYER_H__ */
diff --git a/src/runtime/NEON/functions/NELSTMLayer.cpp b/src/runtime/NEON/functions/NELSTMLayer.cpp
index 42b805794b..ee2b2f4b28 100644
--- a/src/runtime/NEON/functions/NELSTMLayer.cpp
+++ b/src/runtime/NEON/functions/NELSTMLayer.cpp
@@ -38,15 +38,18 @@ using namespace arm_compute;
 using namespace arm_compute::misc::shape_calculator;
 
 NELSTMLayer::NELSTMLayer(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(),
-      _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)
+    : _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(),
+      _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(),
+      _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)
 {
 }
 
@@ -65,6 +68,8 @@ void NELSTMLayer::configure(const ITensor *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())
@@ -117,7 +122,7 @@ void NELSTMLayer::configure(const ITensor *input,
     _concat_weights_forget_gate.configure(weights_vector, &_forget_gate_out6, Window::DimX);
 
     _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();
@@ -130,7 +135,7 @@ void NELSTMLayer::configure(const ITensor *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_ZERO);
-        _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;
@@ -139,6 +144,20 @@ void NELSTMLayer::configure(const ITensor *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_ZERO);
+        // 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(&_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
@@ -170,7 +189,7 @@ void NELSTMLayer::configure(const ITensor *input,
         _memory_group.manage(&_input_gate_out1);
         _memory_group.manage(&_input_gate_out4);
 
-        _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;
 
@@ -178,7 +197,7 @@ void NELSTMLayer::configure(const ITensor *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_ZERO);
-            _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;
@@ -187,6 +206,21 @@ void NELSTMLayer::configure(const ITensor *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_ZERO);
+            // 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(&_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));
     }
 
@@ -200,7 +234,7 @@ void NELSTMLayer::configure(const ITensor *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);
@@ -208,10 +242,25 @@ void NELSTMLayer::configure(const ITensor *input,
     _cell_state_out2.allocator()->allocate();
     _memory_group.manage(&_cell_state_out4);
     _accum_cell_state1.configure(&_cell_state_out1, &_cell_state_out3, &_cell_state_out4, ConvertPolicy::SATURATE);
-    _activation_cell_state.configure(&_cell_state_out4, nullptr, activation_info);
+    Tensor *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_ZERO);
+        // 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(&_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_ZERO);
-    _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_ZERO);
+    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_ZERO);
     _accum_cell_state2.configure(&_cell_state_out5, &_cell_state_out3, &_cell_state_out1, ConvertPolicy::SATURATE);
     _cell_state_out3.allocator()->allocate();
@@ -238,7 +287,7 @@ void NELSTMLayer::configure(const ITensor *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();
@@ -250,7 +299,7 @@ void NELSTMLayer::configure(const ITensor *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_ZERO);
-        _accum_output2.configure(&_output4, &_output3, &_output1, ConvertPolicy::SATURATE);
+        _accum_output1.configure(&_output4, &_output3, &_output1, ConvertPolicy::SATURATE);
         _output4.allocator()->allocate();
         output_gate_out = &_output1;
 
@@ -261,6 +310,20 @@ void NELSTMLayer::configure(const ITensor *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_ZERO);
+        // 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(&_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
@@ -362,6 +425,31 @@ Status NELSTMLayer::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())
     {
@@ -388,13 +476,20 @@ Status NELSTMLayer::validate(const ITensorInfo *input,
     ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(inputs_vector, &forget_gate_concat, Window::DimX));
 
     // Validate forget gate
-    ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, input_to_forget_weights, forget_gate_bias, &forget_gate));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, input_to_forget_weights, (lstm_params.use_layer_norm()) ? nullptr : forget_gate_bias, &forget_gate));
 
     if(lstm_params.has_peephole_opt())
     {
         ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
         ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE));
     }
+    if(lstm_params.use_layer_norm())
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(NEMeanStdDevNormalizationLayer::validate(&forget_gate));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&forget_gate, lstm_params.forget_layer_norm_weights(), &forget_gate, 1, ConvertPolicy::SATURATE,
+                                                                              RoundingPolicy::TO_ZERO));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE));
+    }
     ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayerKernel::validate(&forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
 
     // Validate input gate
@@ -413,7 +508,7 @@ Status NELSTMLayer::validate(const ITensorInfo *input,
         TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0);
         TensorInfo  lstm_gate_concat          = TensorInfo(lstm_weights_concat_shape, 1, input->data_type());
         ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(lstm_weights, &lstm_gate_concat, Window::DimX));
-        ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), lstm_params.input_gate_bias(), &input_gate));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::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())
         {
@@ -422,6 +517,13 @@ Status NELSTMLayer::validate(const ITensorInfo *input,
             ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
             ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE));
         }
+
+        if(lstm_params.use_layer_norm())
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(NEMeanStdDevNormalizationLayer::validate(&input_gate));
+            ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&input_gate, lstm_params.input_layer_norm_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
+            ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&input_gate, lstm_params.input_gate_bias(), &input_gate, ConvertPolicy::SATURATE));
+        }
         ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayerKernel::validate(&input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
     }
     else
@@ -430,9 +532,16 @@ Status NELSTMLayer::validate(const ITensorInfo *input,
     }
 
     // Validate cell state
-    ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, input_to_cell_weights, cell_bias, &cell_state_tmp));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, input_to_cell_weights, (lstm_params.use_layer_norm()) ? nullptr : cell_bias, &cell_state_tmp));
     ARM_COMPUTE_RETURN_ON_ERROR(NEGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &cell_state_tmp, 1.f, 0.f, GEMMInfo()));
     ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
+    if(lstm_params.use_layer_norm())
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(NEMeanStdDevNormalizationLayer::validate(&cell_state_tmp));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&cell_state_tmp, lstm_params.cell_layer_norm_weights(), &cell_state_tmp, 1, ConvertPolicy::SATURATE,
+                                                                              RoundingPolicy::TO_ZERO));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE));
+    }
     ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayerKernel::validate(&cell_state_tmp, nullptr, activation_info));
     ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
     ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
@@ -451,7 +560,7 @@ Status NELSTMLayer::validate(const ITensorInfo *input,
     TensorInfo  in_out_gate_concat          = TensorInfo(in_out_weights_concat_shape, 1, input->data_type());
     ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(in_out_weights, &in_out_gate_concat, Window::DimX));
 
-    ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, input_to_output_weights, output_gate_bias, &output_gate_tmp));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, input_to_output_weights, (lstm_params.use_layer_norm()) ? nullptr : output_gate_bias, &output_gate_tmp));
 
     if(lstm_params.has_peephole_opt())
     {
@@ -459,6 +568,13 @@ Status NELSTMLayer::validate(const ITensorInfo *input,
                                                                               RoundingPolicy::TO_ZERO));
         ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE));
     }
+    if(lstm_params.use_layer_norm())
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(NEMeanStdDevNormalizationLayer::validate(&output_gate_tmp));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(&output_gate_tmp, lstm_params.output_layer_norm_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE,
+                                                                              RoundingPolicy::TO_ZERO));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, ConvertPolicy::SATURATE));
+    }
     ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayerKernel::validate(&output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
 
     // Validate output state
@@ -504,7 +620,13 @@ void NELSTMLayer::run()
     if(_run_peephole_opt)
     {
         NEScheduler::get().schedule(&_pixelwise_mul_forget_gate, Window::DimY);
-        _accum_forget_gate2.run();
+        _accum_forget_gate1.run();
+    }
+    if(_is_layer_norm_lstm)
+    {
+        _mean_std_norm_forget_gate.run();
+        NEScheduler::get().schedule(&_pixelwise_mul_forget_gate_coeff, Window::DimY);
+        NEScheduler::get().schedule(&_accum_forget_gate_bias, Window::DimY);
     }
     NEScheduler::get().schedule(&_activation_forget_gate, Window::DimY);
 
@@ -527,7 +649,14 @@ void NELSTMLayer::run()
         if(_run_peephole_opt)
         {
             NEScheduler::get().schedule(&_pixelwise_mul_input_gate, Window::DimY);
-            _accum_input_gate2.run();
+            _accum_input_gate1.run();
+        }
+
+        if(_is_layer_norm_lstm)
+        {
+            _mean_std_norm_input_gate.run();
+            NEScheduler::get().schedule(&_pixelwise_mul_input_gate_coeff, Window::DimY);
+            NEScheduler::get().schedule(&_accum_input_gate_bias, Window::DimY);
         }
         NEScheduler::get().schedule(&_activation_input_gate, Window::DimY);
     }
@@ -536,6 +665,12 @@ void NELSTMLayer::run()
     NEScheduler::get().schedule(&_transpose_cell_state, Window::DimY);
     _gemm_cell_state1.run();
     NEScheduler::get().schedule(&_accum_cell_state1, Window::DimY);
+    if(_is_layer_norm_lstm)
+    {
+        _mean_std_norm_cell_gate.run();
+        NEScheduler::get().schedule(&_pixelwise_mul_cell_gate_coeff, Window::DimY);
+        NEScheduler::get().schedule(&_accum_cell_gate_bias, Window::DimY);
+    }
     NEScheduler::get().schedule(&_activation_cell_state, Window::DimY);
     NEScheduler::get().schedule(&_pixelwise_mul_cell_state1, Window::DimY);
     NEScheduler::get().schedule(&_pixelwise_mul_cell_state2, Window::DimY);
@@ -550,7 +685,13 @@ void NELSTMLayer::run()
     if(_run_peephole_opt)
     {
         NEScheduler::get().schedule(&_pixelwise_mul_output_state1, Window::DimY);
-        _accum_output2.run();
+        _accum_output1.run();
+    }
+    if(_is_layer_norm_lstm)
+    {
+        _mean_std_norm_output_gate.run();
+        NEScheduler::get().schedule(&_pixelwise_mul_output_gate_coeff, Window::DimY);
+        NEScheduler::get().schedule(&_accum_output_gate_bias, Window::DimY);
     }
     NEScheduler::get().schedule(&_activation_output, Window::DimY);
 
diff --git a/tests/validation/NEON/LSTMLayer.cpp b/tests/validation/NEON/LSTMLayer.cpp
index c503972ba9..45beb36e60 100644
--- a/tests/validation/NEON/LSTMLayer.cpp
+++ b/tests/validation/NEON/LSTMLayer.cpp
@@ -157,7 +157,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NELSTMLayerFixture<float>, framework::DatasetMo
                                                                                                                  DataType::F32)),
                                                                                                                  framework::dataset::make("ProjectionOpt", { true, false })),
                                                                                                          framework::dataset::make("PeepholeOpt", { true, false })),
-                                                                                                 framework::dataset::make("UseLayerNorm", { false })))
+                                                                                                 framework::dataset::make("UseLayerNorm", { true, false })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f32);
@@ -171,7 +171,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NELSTMLayerFixture<half>, framework::DatasetMod
                                                                                                                         DataType::F16)),
                                                                                                                 framework::dataset::make("ProjectionOpt", { true, false })),
                                                                                                         framework::dataset::make("PeepholeOpt", { true, false })),
-                                                                                                framework::dataset::make("UseLayerNorm", { false })))
+                                                                                                framework::dataset::make("UseLayerNorm", { true, false })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f16);