COMPMID-3237: Implement NEQLSTMLayer

COMPMID-3082: Extend NEQLSTMLayer with enhancements

Change-Id: I88175b7bf69494a4eae510b74176fe8a0d6cd770
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2969
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

7 files changed, 437 insertions, 69 deletions

diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h
index c6b40eb30e..8f47c5089d 100644
--- a/arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -63,14 +63,14 @@ public:
      * kernels change the layout of the original matrices to be more cache-friendly.
      *
      * @param[in]  input0 Input tensor containing the interleaved Matrix A. Data type supported: U8/QASYMM8/S8/QASYMM8_SIGNED
-     * @param[in]  input1 Input tensor containing the transposed1xW Matrix B. Data type supported: U8/QASYMM8/S8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
+     * @param[in]  input1 Input tensor containing the transposed1xW Matrix B. Data type supported: U8/QASYMM8/S8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL
      * @param[out] output Output tensor to store the result of matrix multiplication. Data type supported: S32
      */
     void configure(const ITensor *input0, const ITensor *input1, ITensor *output);
     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpMatrixMultiplyKernel
      *
      * @param[in] input0 Input tensor info containing the interleaved Matrix A. Data type supported: U8/QASYMM8/S8/QASYMM8_SIGNED
-     * @param[in] input1 Input tensor info containing the transposed Matrix B. Data type supported: U8/QASYMM8/S8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
+     * @param[in] input1 Input tensor info containing the transposed Matrix B. Data type supported: U8/QASYMM8/S8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL
      * @param[in] output Output tensor info to store the result of matrix multiplication. Data type supported: S32
      *
      * @return a status
diff --git a/arm_compute/core/utils/misc/InfoHelpers.h b/arm_compute/core/utils/misc/InfoHelpers.h
index b572de2433..8cf701c124 100644
--- a/arm_compute/core/utils/misc/InfoHelpers.h
+++ b/arm_compute/core/utils/misc/InfoHelpers.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,6 +26,7 @@
 
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/common/LSTMParams.h"
 
 namespace arm_compute
 {
@@ -58,6 +59,38 @@ inline bool is_relu6(ActivationLayerInfo activation_info)
                                  && activation_info.a() == 6.f;
     return activation_info.enabled() && (is_lu_bounded_relu || is_bounded_relu);
 }
+
+/** Build LSTMParams<ITensorInfo> object by extracting the metadata from each
+ * tensor.
+ *
+ * @param[in]  lstm_params      The LSTMParams<T> object containing the tensors.
+ * @param[out] lstm_params_info The LSTMParams<ITensorInfo> to be constructed.
+ *
+ */
+template <typename T>
+inline void build_lstm_params_tensor_info(const LSTMParams<T>     &lstm_params,
+                                          LSTMParams<ITensorInfo> *lstm_params_info)
+{
+    if(lstm_params.has_peephole_opt())
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_output_weights());
+        lstm_params_info->set_peephole_params(lstm_params.cell_to_forget_weights()->info(), lstm_params.cell_to_output_weights()->info());
+    }
+    if(lstm_params.has_projection())
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(lstm_params.projection_weights());
+        lstm_params_info->set_projection_params(lstm_params.projection_weights()->info(),
+                                                lstm_params.projection_bias() != nullptr ? lstm_params.projection_bias()->info() : nullptr);
+    }
+    if(!lstm_params.has_cifg_opt())
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights(), lstm_params.input_gate_bias());
+
+        const ITensorInfo *cell_to_input_weights_info = (lstm_params.has_peephole_opt()) ? lstm_params.cell_to_input_weights()->info() : nullptr;
+        lstm_params_info->set_cifg_params(lstm_params.input_to_input_weights()->info(), lstm_params.recurrent_to_input_weights()->info(),
+                                          cell_to_input_weights_info, lstm_params.input_gate_bias()->info());
+    }
+}
 } // namespace info_helpers
 } // namespace utils
 } // namespace arm_compute
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index abad8d482e..de364fa9af 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2019 ARM Limited.
+ * Copyright (c) 2016-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -120,6 +120,7 @@
 #include "arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h"
 #include "arm_compute/runtime/NEON/functions/NEPoolingLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEPriorBoxLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEQLSTMLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEQuantizationLayer.h"
 #include "arm_compute/runtime/NEON/functions/NERNNLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEROIAlignLayer.h"
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
index 74dedcf4c5..11683c5b95 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
@@ -84,7 +84,7 @@ public:
      * @note The @p output type is S32 if @p gemm_info.type == GEMMLowpOutputStageType::NONE. It is QASYMM8/QASYMM8_SIGNED otherwise
      *
      * @param[in]  a         First input tensor  (Matrix A). Data type supported: QASYMM8/QASYMM8_SIGNED.
-     * @param[in]  b         Second input tensor (Matrix B). Data type supported: same as @p a
+     * @param[in]  b         Second input tensor (Matrix B). Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL.
      * @param[in]  c         Third input tensor  (Matrix C). It can be a nullptr. Data type supported: S32
      * @param[out] output    Output tensor. Data type supported: Data type supported: S32/QASYMM8/QASYMM8_SIGNED
      * @param[in]  gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
@@ -96,7 +96,7 @@ public:
      * @note The @p output type is S32 if @p gemm_info.type == GEMMLowpOutputStageType::NONE. It is QASYMM8/QASYMM8_SIGNED otherwise
      *
      * @param[in] a         First input tensor info  (Matrix A). Data type supported: QASYMM8/QASYMM8_SIGNED.
-     * @param[in] b         Second input tensor info (Matrix B). Data type supported: same as @p a
+     * @param[in] b         Second input tensor info (Matrix B). Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL.
      * @param[in] c         Third input tensor  info (Matrix C). It can be a nullptr. Data type supported: S32
      * @param[in] output    Output tensor info. Data type supported: Data type supported: S32/QASYMM8/QASYMM8_SIGNED
      * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
diff --git a/arm_compute/runtime/NEON/functions/NELSTMLayer.h b/arm_compute/runtime/NEON/functions/NELSTMLayer.h
index ae13d0c36f..e85e87b88e 100644
--- a/arm_compute/runtime/NEON/functions/NELSTMLayer.h
+++ b/arm_compute/runtime/NEON/functions/NELSTMLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -68,22 +68,23 @@ public:
      * @param[out] cell_state_out              2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
      * @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_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]  lstm_params                 Weights tensors used in peephole optimization:
+     *                                         input_to_input_weights     (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
+     *                                         recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     *                                         cell_to_input_weights      (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
+     *                                         cell_to_forget_weights     (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         cell_to_output_weights     (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         input_gate_bias            (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
+     *                                         projection_weights         (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     *                                         projection_bias            (Optional) 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
+     *                                         input_layer_norm_weights   (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         forget_layer_norm_weights  (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         cell_layer_norm_weights    (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                         output_layer_norm_weights  (Optional) 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]  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.
      */
     void configure(const ITensor *input,
                    const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights,
@@ -112,22 +113,23 @@ public:
      * @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.
-     *                                        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] lstm_params                 Weights tensors used in peephole optimization:
+     *                                        input_to_input_weights     (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: Same as @p input.
+     *                                        recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     *                                        cell_to_input_weights      (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input.
+     *                                        cell_to_forget_weights     (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        cell_to_output_weights     (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        input_gate_bias            (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input
+     *                                        projection_weights         (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input.
+     *                                        projection_bias            (Optional) 1D weights tensor with dimensions [output_size]. Data type supported: Same as @p input.
+     *                                        input_layer_norm_weights   (Optional) 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        forget_layer_norm_weights  (Optional) 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        cell_layer_norm_weights    (Optional) 1D weights tensor info with dimensions [num_units]. Data type supported: Same as @p input.
+     *                                        output_layer_norm_weights  (Optional) 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] 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.
      *
      * @return a status
      */
diff --git a/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h b/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h
new file mode 100644
index 0000000000..a37909b775
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h
@@ -0,0 +1,332 @@
+/*
+ * Copyright (c) 2020 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_NEQLSTMLAYER_H
+#define ARM_COMPUTE_NEQLSTMLAYER_H
+
+#include "arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h"
+#include "arm_compute/core/NEON/kernels/NEArithmeticSubtractionKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h"
+#include "arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEActivationLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
+#include "arm_compute/runtime/NEON/functions/NETranspose.h"
+
+#include "arm_compute/runtime/common/LSTMParams.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ITensor;
+
+/** Basic function to run @ref NEQLSTMLayer
+ *
+ * This function calls the following NEON functions/kernels:
+ *
+ * -# @ref NEActivationLayer                                     Activation functions (tanh and logistic)
+ * -# @ref NEArithmeticAdditionKernel                            Elementwise addition
+ * -# @ref NEArithmeticSubtractionKernel                         Elementwise subtraction
+ * -# @ref NEGEMMLowpMatrixMultiplyCore                          Quantized matrix multiplication core. Accumulators are 32-bit integers
+ * -# @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint   Convert 32-bit integers into QSYMM16
+ * -# @ref NEGEMMLowpMatrixAReductionKernel                      For precomputing effective biases to use
+ * -# @ref NEPixelWiseMultiplicationKernel                       Elementwise multiplication
+ * -# @ref NETranspose                                           Transpose function for reshaping the weights
+ * */
+class NEQLSTMLayer : public IFunction
+{
+public:
+    /** Default constructor */
+    NEQLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEQLSTMLayer(const NEQLSTMLayer &) = delete;
+    /** Default move constructor */
+    NEQLSTMLayer(NEQLSTMLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEQLSTMLayer &operator=(const NEQLSTMLayer &) = delete;
+    /** Default move assignment operator */
+    NEQLSTMLayer &operator=(NEQLSTMLayer &&) = default;
+    /** Initialize function's tensors.
+     *
+     * @param[in]  input                       Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: QASYMM8_SIGNED.
+     * @param[in]  input_to_forget_weights     2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  input_to_cell_weights       2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  input_to_output_weights     2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  recurrent_to_cell_weights   2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  forget_gate_bias            1D weights tensor with dimensions [num_units]. Data type supported: S32.
+     * @param[in]  cell_bias                   1D weights tensor with dimensions [num_units]. Data type supported: S32.
+     * @param[in]  output_gate_bias            1D weights tensor with dimensions [num_units]. Data type supported: S32.
+     * @param[in]  cell_state_in               2D tensor with dimensions [output_size, batch_size]. Data type supported:  QSYMM16.
+     * @param[in]  output_state_in             2D tensor with dimensions [num_units, batch_size]. Data type supported: Same as @p input.
+     * @param[out] cell_state_out              Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size]. Data type supported:  QSYMM16.
+     * @param[out] output_state_out            Destination tensor. Output is a 2D tensor with dimensions [num_units, batch_size].Data types supported: Same as @p input.
+     * @param[in]  lstm_params                 Weights tensors used in peephole, CIFG and layer normalization optimizations:
+     *                                         input_intermediate_scale   Scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate.
+     *                                         forget_intermediate_scale  Scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate.
+     *                                         cell_intermediate_scale    Scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate.
+     *                                         output_intermediate_scale  Scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate.
+     *                                         hidden_state_zero          The zero point of the hidden state.
+     *                                         hidden_state_scale         The scale of the hidden state.
+     *                                         input_to_input_weights     (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
+     *                                         recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     *                                         cell_to_input_weights      (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: QSYMM16.
+     *                                         cell_to_forget_weights     (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         cell_to_output_weights     (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         input_gate_bias            (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: S32.
+     *                                         projection_weights         (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     *                                         projection_bias            (Optional) 1D weights tensor with dimensions [output_size]. S32.
+     *                                         input_layer_norm_weights   (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         forget_layer_norm_weights  (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         cell_layer_norm_weights    (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         output_layer_norm_weights  (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         cell_threshold             (Optional) 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.
+     *                                         projection_threshold       (Optional) 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.
+     */
+    void configure(const ITensor *input,
+                   const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights,
+                   const ITensor *recurrent_to_forget_weights, const ITensor *recurrent_to_cell_weights, const ITensor *recurrent_to_output_weights,
+                   const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias,
+                   const ITensor *cell_state_in, const ITensor *output_state_in,
+                   ITensor *cell_state_out, ITensor *output_state_out,
+                   const LSTMParams<ITensor> &lstm_params);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref NEQLSTMLayer
+     *
+     * @param[in]  input                       Source tensor info. Input is a 2D tensor info with dimensions [input_size, batch_size]. Data types supported: QASYMM8_SIGNED.
+     * @param[in]  input_to_forget_weights     2D weights tensor info with dimensions [input_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  input_to_cell_weights       2D weights tensor info with dimensions [input_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  input_to_output_weights     2D weights tensor info with dimensions [input_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  recurrent_to_cell_weights   2D weights tensor info with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     * @param[in]  forget_gate_bias            1D weights tensor info with dimensions [num_units]. Data type supported: S32.
+     * @param[in]  cell_bias                   1D weights tensor info with dimensions [num_units]. Data type supported: S32.
+     * @param[in]  output_gate_bias            1D weights tensor info with dimensions [num_units]. Data type supported: S32.
+     * @param[in]  cell_state_in               2D tensor info with dimensions [num_units, batch_size]. Data type supported:  QSYMM16.
+     * @param[in]  output_state_in             2D tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
+     * @param[out] cell_state_out              Destination tensor info. Output is a 2D tensor info with dimensions [num_units, batch_size]. Data type supported:  QSYMM16.
+     * @param[out] output_state_out            Destination tensor info. Output is a 2D tensor info with dimensions [output_size, batch_size].Data types supported: Same as @p input.
+     * @param[in]  lstm_params                 Weights tensors info used in peephole, CIFG and layer normalization optimizations:
+     *                                         input_intermediate_scale   Scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate.
+     *                                         forget_intermediate_scale  Scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate.
+     *                                         cell_intermediate_scale    Scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate.
+     *                                         output_intermediate_scale  Scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate.
+     *                                         hidden_state_zero          The zero point of the hidden state.
+     *                                         hidden_state_scale         The scale of the hidden state.
+     *                                         input_to_input_weights     (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
+     *                                         recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     *                                         cell_to_input_weights      (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: QSYMM16.
+     *                                         cell_to_forget_weights     (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         cell_to_output_weights     (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         input_gate_bias            (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: S32.
+     *                                         projection_weights         (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
+     *                                         projection_bias            (Optional) 1D weights tensor with dimensions [output_size]. S32.
+     *                                         input_layer_norm_weights   (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         forget_layer_norm_weights  (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         cell_layer_norm_weights    (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         output_layer_norm_weights  (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
+     *                                         cell_threshold             (Optional) 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.
+     *                                         projection_threshold       (Optional) 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.
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input,
+                           const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
+                           const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
+                           const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
+                           const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in,
+                           const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out,
+                           const LSTMParams<ITensorInfo> &lstm_params);
+
+    // Inherited methods overridden:
+    void run() override;
+    void prepare() override;
+
+private:
+    /** Internal method to configure matrix multiplication plus output stage of each gate.
+     *
+     * @param[in] mm             Matrix multiplication function to use.
+     * @param[in] outstage       Output stage function to use.
+     * @param[in] gemmlowp_info  GEMMLowp metadata to be used by the output stage.
+     * @param[in] mm_input       Input tensor to matrix multiplication function.
+     * @param[in] mm_weights     Weights tensor to matrix multiplication function.
+     * @param[in] bias           Bias tensor to matrix multiplication function.
+     * @param[in] outstage_res   Tensor to be used for storing the result of the output stage.
+     * @param[in] gemmlowp_scale Real multiplier to be used computing multiplier and shift for requantization.
+     * @param[in] mm_res_info    Tensor info to be used to initialize matrix multiplication result tensor.
+     * @param[in] mm_res_info    Tensor info to be used to initialize output stage result tensor.
+     *
+     */
+    void configure_mm(NEGEMMLowpMatrixMultiplyCore &mm, NEGEMMLowpOutputStage &outstage, GEMMLowpOutputStageInfo &gemmlowp_info,
+                      const ITensor *mm_input, const ITensor *mm_weights, const ITensor *bias, Tensor *mm_res,
+                      Tensor *outstage_res, float gemmlowp_scale,
+                      const TensorInfo &mm_res_info, const TensorInfo &outstage_tensor_info);
+
+    MemoryGroup _memory_group{};
+
+    // Functions used
+    NETranspose                      _transpose_input_to_forget_weights{};
+    NETranspose                      _transpose_input_to_cell_weights{};
+    NETranspose                      _transpose_input_to_output_weights{};
+    NETranspose                      _transpose_input_to_input_weights{};
+    NETranspose                      _transpose_recurrent_to_forget_weights{};
+    NETranspose                      _transpose_recurrent_to_cell_weights{};
+    NETranspose                      _transpose_recurrent_to_output_weights{};
+    NETranspose                      _transpose_recurrent_to_input_weights{};
+    NETranspose                      _transpose_projection_weights{};
+    NEGEMMLowpMatrixAReductionKernel _input_to_input_reduction{};
+    NEGEMMLowpMatrixAReductionKernel _recurrent_to_input_reduction{};
+    NEGEMMLowpMatrixAReductionKernel _input_to_forget_reduction{};
+    NEGEMMLowpMatrixAReductionKernel _recurrent_to_forget_reduction{};
+    NEGEMMLowpMatrixAReductionKernel _input_to_cell_reduction{};
+    NEGEMMLowpMatrixAReductionKernel _recurrent_to_cell_reduction{};
+    NEGEMMLowpMatrixAReductionKernel _input_to_output_reduction{};
+    NEGEMMLowpMatrixAReductionKernel _recurrent_to_output_reduction{};
+    NEGEMMLowpMatrixAReductionKernel _projection_reduction{};
+    NEArithmeticAdditionKernel       _projection_bias_add{};
+    NEGEMMLowpMatrixMultiplyCore     _mm_input_to_forget{};
+    NEGEMMLowpMatrixMultiplyCore     _mm_recurrent_to_forget{};
+    NEPixelWiseMultiplicationKernel  _pixelwise_mul_cell_to_forget{};
+    NEGEMMLowpOutputStage            _input_to_forget_outstage{};
+    NEGEMMLowpOutputStage            _recurrent_to_forget_outstage{};
+    NEGEMMLowpOutputStage            _cell_to_forget_outstage{};
+    NEArithmeticAdditionKernel       _accumulate_input_recurrent_forget{};
+    NEArithmeticAdditionKernel       _accumulate_cell_forget{};
+    NEActivationLayer                _forget_gate_sigmoid{};
+    NEGEMMLowpMatrixMultiplyCore     _mm_input_to_cell{};
+    NEGEMMLowpOutputStage            _input_to_cell_outstage{};
+    NEGEMMLowpMatrixMultiplyCore     _mm_recurrent_to_cell{};
+    NEGEMMLowpOutputStage            _recurrent_to_cell_outstage{};
+    NEArithmeticAdditionKernel       _accumulate_input_recurrent_modulation{};
+    NEActivationLayer                _cell_gate_tanh{};
+    NEArithmeticSubtractionKernel    _input_gate_sub{};
+    NEGEMMLowpMatrixMultiplyCore     _mm_input_to_input{};
+    NEGEMMLowpOutputStage            _input_to_input_outstage{};
+    NEGEMMLowpMatrixMultiplyCore     _mm_recurrent_to_input{};
+    NEGEMMLowpOutputStage            _recurrent_to_input_outstage{};
+    NEArithmeticAdditionKernel       _accumulate_input_recurrent_input{};
+    NEPixelWiseMultiplicationKernel  _pixelwise_mul_cell_to_input{};
+    NEGEMMLowpOutputStage            _cell_to_input_outstage{};
+    NEArithmeticAdditionKernel       _accumulate_cell_input{};
+    NEActivationLayer                _input_gate_tanh{};
+    NEPixelWiseMultiplicationKernel  _pixelwise_mul_forget_cell{};
+    NEPixelWiseMultiplicationKernel  _pixelwise_mul_input_cell{};
+    NEArithmeticAdditionKernel       _add_forget_cell{};
+    NEActivationLayer                _cell_clip{};
+    NEGEMMLowpMatrixMultiplyCore     _mm_input_to_output{};
+    NEGEMMLowpOutputStage            _input_to_output_outstage{};
+    NEGEMMLowpMatrixMultiplyCore     _mm_recurrent_to_output{};
+    NEGEMMLowpOutputStage            _recurrent_to_output_outstage{};
+    NEArithmeticAdditionKernel       _accumulate_input_recurrent_output{};
+    NEPixelWiseMultiplicationKernel  _pixelwise_mul_cell_to_output{};
+    NEArithmeticAdditionKernel       _accumulate_cell_to_output{};
+    NEActivationLayer                _output_gate_sigmoid{};
+    NEActivationLayer                _hidden_tanh{};
+    NEPixelWiseMultiplicationKernel  _pixelwise_mul_hidden{};
+    NEGEMMLowpOutputStage            _hidden_outstage{};
+    NEGEMMLowpMatrixMultiplyCore     _mm_projection{};
+    NEGEMMLowpOutputStage            _projection_outstage{};
+    NEArithmeticAdditionKernel       _accumulate_projection{};
+    NEActivationLayer                _projection_clip{};
+
+    // Tensor pointers
+    const ITensor *_input_to_input_weights
+    {
+        nullptr
+    };
+    const ITensor *_recurrent_to_input_weights{ nullptr };
+    const ITensor *_projection_bias{ nullptr };
+    const ITensor *_input_to_forget_weights{ nullptr };
+    const ITensor *_input_to_cell_weights{ nullptr };
+    const ITensor *_input_to_output_weights{ nullptr };
+    const ITensor *_recurrent_to_forget_weights{ nullptr };
+    const ITensor *_recurrent_to_cell_weights{ nullptr };
+    const ITensor *_recurrent_to_output_weights{ nullptr };
+    const ITensor *_projection_weights{ nullptr };
+
+    // Temporary tensors
+    Tensor _input_to_forget_weights_transposed{ nullptr };
+    Tensor _input_to_cell_weights_transposed{ nullptr };
+    Tensor _input_to_output_weights_transposed{ nullptr };
+    Tensor _input_to_input_weights_transposed{ nullptr };
+    Tensor _recurrent_to_forget_weights_transposed{ nullptr };
+    Tensor _recurrent_to_cell_weights_transposed{ nullptr };
+    Tensor _recurrent_to_output_weights_transposed{ nullptr };
+    Tensor _recurrent_to_input_weights_transposed{ nullptr };
+    Tensor _projection_weights_transposed{ nullptr };
+    Tensor _input_to_input_eff_bias{ nullptr };
+    Tensor _recurrent_to_input_eff_bias{ nullptr };
+    Tensor _input_to_forget_eff_bias{ nullptr };
+    Tensor _recurrent_to_forget_eff_bias{ nullptr };
+    Tensor _input_to_cell_eff_bias{ nullptr };
+    Tensor _recurrent_to_cell_eff_bias{ nullptr };
+    Tensor _input_to_output_eff_bias{ nullptr };
+    Tensor _recurrent_to_output_eff_bias{ nullptr };
+    Tensor _projection_reduction_res{ nullptr };
+    Tensor _projection_eff_bias{ nullptr };
+    Tensor _mm_input_to_forget_res{ nullptr };
+    Tensor _mm_recurrent_to_forget_res{ nullptr };
+    Tensor _mul_cell_to_forget_res{ nullptr };
+    Tensor _input_to_forget_outstage_res{ nullptr };
+    Tensor _cell_to_forget_outstage_res{ nullptr };
+    Tensor _recurrent_to_forget_outstage_res{ nullptr };
+    Tensor _forget_gate{ nullptr };
+    Tensor _mm_input_to_cell_res{ nullptr };
+    Tensor _input_to_cell_outstage_res{ nullptr };
+    Tensor _mm_recurrent_to_cell_res{ nullptr };
+    Tensor _recurrent_to_cell_outstage_res{ nullptr };
+    Tensor _cell_gate{ nullptr };
+    Tensor _mul_input_cell_res{ nullptr };
+    Tensor _mm_input_to_input_res{ nullptr };
+    Tensor _input_to_input_outstage_res{ nullptr };
+    Tensor _mm_recurrent_to_input_res{ nullptr };
+    Tensor _mul_cell_to_input_res{ nullptr };
+    Tensor _cell_to_input_outstage_res{ nullptr };
+    Tensor _recurrent_to_input_outstage_res{ nullptr };
+    Tensor _input_gate{ nullptr };
+    Tensor _mm_input_to_output_res{ nullptr };
+    Tensor _input_to_output_outstage_res{ nullptr };
+    Tensor _mm_recurrent_to_output_res{ nullptr };
+    Tensor _mul_cell_to_output_res{ nullptr };
+    Tensor _recurrent_to_output_outstage_res{ nullptr };
+    Tensor _output_gate{ nullptr };
+    Tensor _hidden_mul_res{ nullptr };
+    Tensor _mm_projection_res{ nullptr };
+    Tensor _projection_outstage_res{ nullptr };
+    Tensor _ones{ nullptr };
+
+    bool _is_prepared{ false };
+    bool _has_cifg{ false };
+    bool _has_cell_clipping{ false };
+    bool _has_projection{ false };
+    bool _has_projection_clipping{ false };
+    bool _has_peephole{ false };
+};
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_NEQLSTMLAYER_H */
diff --git a/arm_compute/runtime/common/LSTMParams.h b/arm_compute/runtime/common/LSTMParams.h
index f16945730e..e21ddd7af1 100644
--- a/arm_compute/runtime/common/LSTMParams.h
+++ b/arm_compute/runtime/common/LSTMParams.h
@@ -54,10 +54,10 @@ public:
           _output_layer_norm_weights(nullptr),
           _cell_clip(0.f),
           _projection_clip(0.0f),
-          _input_gate_matmul_scale(0.0f),
-          _forget_gate_matmul_scale(0.0f),
-          _cell_gate_matmul_scale(0.0f),
-          _output_gate_matmul_scale(0.0f),
+          _input_intermediate_scale(0.0f),
+          _forget_intermediate_scale(0.0f),
+          _cell_intermediate_scale(0.0f),
+          _output_intermediate_scale(0.0f),
           _hidden_state_zero(0.0f),
           _hidden_state_scale(0),
           _has_peephole_opt(false),
@@ -74,10 +74,10 @@ public:
     ~LSTMParams() = default;
     /** Set CIFG tensor parameters.
      *
-     * @param[in] input_to_input_weights     2D weights tensor with dimensions [input_size, num_units]. Data types supported: F16/F32.
+     * @param[in] input_to_input_weights     2D weights tensor with dimensions [input_size, num_units]. Data types supported: QSYMM8/F16/F32.
      * @param[in] recurrent_to_input_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Same as @p input_to_input_weights.
      * @param[in] cell_to_input_weights      1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: Same as @p input_to_input_weights.
-     * @param[in] input_gate_bias            1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input_to_input_weights
+     * @param[in] input_gate_bias            1D weights tensor with dimensions [num_units]. Data type supported: Same as @p input_to_input_weights, S32 when @p input_to_input_weights is QSYMM8
      *
      * @return Reference to this LSTMParams object
      */
@@ -92,8 +92,8 @@ public:
     }
     /** Set projection tensor parameters.
      *
-     * @param[in] projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Data types supported: F16/F32.
-     * @param[in] projection_bias    1D weights tensor with dimensions [output_size]. Data type supported: Same as @p projection_weights.
+     * @param[in] projection_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: Data types supported: QSYMM8/F16/F32.
+     * @param[in] projection_bias    1D weights tensor with dimensions [output_size]. Data type supported: Same as @p projection_weights, S32 when @p input_to_input_weights is QSYMM8.
      *
      * @return Reference to this LSTMParams object
      */
@@ -106,8 +106,8 @@ public:
     }
     /** Set peephole tensor parameters.
      *
-     * @param[in] cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Data types supported: F16/F32.
-     * @param[in] cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p cell_to_input_weights.
+     * @param[in] cell_to_forget_weights 1D weights tensor with dimensions [num_units]. Data type supported: Data types supported: QSYMM16/F16/F32.
+     * @param[in] cell_to_output_weights 1D weights tensor with dimensions [num_units]. Data type supported: Same as @p cell_to_forget_weights.
      *
      * @return Reference to this LSTMParams object
      */
@@ -120,7 +120,7 @@ public:
     }
     /** 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] input_layer_norm_weights  1D weights tensor with dimensions [num_units]. Data type supported: Data types supported: QSYMM16/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.
@@ -164,19 +164,19 @@ public:
 
     /** Set scale of the intermediate results of matmul of each layer parameters.
      *
-     * @param[in] input_gate_matmul_scale  Scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate.
-     * @param[in] forget_gate_matmul_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate.
-     * @param[in] cell_gate_matmul_scale   Scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate.
-     * @param[in] output_gate_matmul_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate.
+     * @param[in] input_intermediate_scale  Scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate.
+     * @param[in] forget_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate.
+     * @param[in] cell_intermediate_scale   Scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate.
+     * @param[in] output_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate.
      *
      * @return Reference to this LSTMParams object
      */
-    LSTMParams &set_matmul_scale_params(float input_gate_matmul_scale, float forget_gate_matmul_scale, float cell_gate_matmul_scale, float output_gate_matmul_scale)
+    LSTMParams &set_matmul_scale_params(float input_intermediate_scale, float forget_intermediate_scale, float cell_intermediate_scale, float output_intermediate_scale)
     {
-        _input_gate_matmul_scale  = input_gate_matmul_scale;
-        _forget_gate_matmul_scale = forget_gate_matmul_scale;
-        _cell_gate_matmul_scale   = cell_gate_matmul_scale;
-        _output_gate_matmul_scale = output_gate_matmul_scale;
+        _input_intermediate_scale  = input_intermediate_scale;
+        _forget_intermediate_scale = forget_intermediate_scale;
+        _cell_intermediate_scale   = cell_intermediate_scale;
+        _output_intermediate_scale = output_intermediate_scale;
         return *this;
     }
 
@@ -187,7 +187,7 @@ public:
      *
      * @return Reference to this LSTMParams object
      */
-    LSTMParams &set_matmul_scale_params(int32_t hidden_state_zero, float hidden_state_scale)
+    LSTMParams &set_hidden_state_params(int32_t hidden_state_zero, float hidden_state_scale)
     {
         _hidden_state_zero  = hidden_state_zero;
         _hidden_state_scale = hidden_state_scale;
@@ -264,24 +264,24 @@ public:
         return _projection_clip;
     }
 
-    float input_gate_matmul_scale() const
+    float input_intermediate_scale() const
     {
-        return _input_gate_matmul_scale;
+        return _input_intermediate_scale;
     }
 
-    float forget_gate_matmul_scale() const
+    float forget_intermediate_scale() const
     {
-        return _forget_gate_matmul_scale;
+        return _forget_intermediate_scale;
     }
 
-    float cell_gate_matmul_scale() const
+    float cell_intermediate_scale() const
     {
-        return _cell_gate_matmul_scale;
+        return _cell_intermediate_scale;
     }
 
-    float output_gate_matmul_scale() const
+    float output_intermediate_scale() const
     {
-        return _output_gate_matmul_scale;
+        return _output_intermediate_scale;
     }
 
     int32_t hidden_state_zero() const
@@ -329,10 +329,10 @@ private:
     const T *_output_layer_norm_weights;
     float    _cell_clip;
     float    _projection_clip;
-    float    _input_gate_matmul_scale;
-    float    _forget_gate_matmul_scale;
-    float    _cell_gate_matmul_scale;
-    float    _output_gate_matmul_scale;
+    float    _input_intermediate_scale;
+    float    _forget_intermediate_scale;
+    float    _cell_intermediate_scale;
+    float    _output_intermediate_scale;
     float    _hidden_state_zero;
     int32_t  _hidden_state_scale;
     bool     _has_peephole_opt;