COMPMID-2236: QUANTIZED_16BIT_LSTM operator for NEON

Change-Id: I554023508e09b790ecc1bbdada529697d6c7b616
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1551
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

4 files changed, 447 insertions, 4 deletions

diff --git a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
index bf0a2ca7bf..d11f04a82f 100644
--- a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
@@ -28,6 +28,7 @@
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/NEON/NEAsymm.h"
+#include "arm_compute/core/NEON/NESymm.h"
 #include "arm_compute/core/NEON/wrapper/wrapper.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
@@ -42,7 +43,7 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QSYMM8);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QSYMM8, DataType::QSYMM16);
 
     if(output->tensor_shape().total_size() > 0)
     {
@@ -95,6 +96,27 @@ inline void store_result<float16_t>(float16_t *ptr, const float32x4x4_t &v)
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
 template <typename T>
+inline void store_result(T *ptr, const float32x4x2_t &v)
+{
+    ARM_COMPUTE_UNUSED(ptr, v);
+}
+
+template <>
+inline void store_result<float>(float *ptr, const float32x4x2_t &v)
+{
+    wrapper::vstore(ptr, v.val[0]);
+    wrapper::vstore(ptr + 4, v.val[1]);
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+template <>
+inline void store_result<float16_t>(float16_t *ptr, const float32x4x2_t &v)
+{
+    wrapper::vstore(ptr, vcombine_f16(vcvt_f16_f32(v.val[0]), vcvt_f16_f32(v.val[1])));
+}
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
+template <typename T>
 void run_dequantization_qasymm8(const ITensor *input, ITensor *output, const Window &window)
 {
     const UniformQuantizationInfo &qinfo  = input->info()->quantization_info().uniform();
@@ -180,6 +202,48 @@ void run_dequantization_qsymm8(const ITensor *input, ITensor *output, const Wind
 }
 
 template <typename T>
+void run_dequantization_qsymm16(const ITensor *input, ITensor *output, const Window &window)
+{
+    const UniformQuantizationInfo &qinfo = input->info()->quantization_info().uniform();
+    const float                    scale = qinfo.scale;
+
+    const int  window_step_x  = 8;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    // Collapse window and reset first dimension to handle tail calculations manually
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    // Create iterators
+    Iterator in(input, win_collapsed);
+    Iterator out(output, win_collapsed);
+
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto in_ptr  = reinterpret_cast<const int16_t *>(in.ptr());
+        const auto out_ptr = reinterpret_cast<T *>(out.ptr());
+
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            const auto vin  = wrapper::vloadq(in_ptr + x);
+            const auto vdeq = vdequantize_int16(vin, scale);
+
+            store_result<T>(reinterpret_cast<T *>(out_ptr + x), vdeq);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            int16_t val    = *(in_ptr + x);
+            *(out_ptr + x) = static_cast<T>(dequantize_qsymm16(val, scale));
+        }
+    },
+    in, out);
+}
+
+template <typename T>
 void run_dequantization_core(const ITensor *input, ITensor *output, const Window &window)
 {
     switch(input->info()->data_type())
@@ -190,6 +254,9 @@ void run_dequantization_core(const ITensor *input, ITensor *output, const Window
         case DataType::QSYMM8:
             run_dequantization_qsymm8<T>(input, output, window);
             break;
+        case DataType::QSYMM16:
+            run_dequantization_qsymm16<T>(input, output, window);
+            break;
         default:
             ARM_COMPUTE_ERROR("Unsupported data type.");
     }
@@ -244,4 +311,4 @@ void NEDequantizationLayerKernel::run(const Window &window, const ThreadInfo &in
             ARM_COMPUTE_ERROR("Unsupported data type.");
     }
 }
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute
diff --git a/src/core/NEON/kernels/NEStridedSliceKernel.cpp b/src/core/NEON/kernels/NEStridedSliceKernel.cpp
index ece291e0a3..c33e699999 100644
--- a/src/core/NEON/kernels/NEStridedSliceKernel.cpp
+++ b/src/core/NEON/kernels/NEStridedSliceKernel.cpp
@@ -45,7 +45,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
                                                          DataType::U8, DataType::S8, DataType::QASYMM8,
-                                                         DataType::U16, DataType::S16,
+                                                         DataType::U16, DataType::S16, DataType::QSYMM16,
                                                          DataType::U32, DataType::S32,
                                                          DataType::F16, DataType::F32);
 
diff --git a/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp
index 28f655c529..7b1ad9c2e8 100644
--- a/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEWidthConcatenateLayerKernel.cpp
@@ -61,7 +61,7 @@ Status validate_arguments(const ITensorInfo *input, unsigned int width_offset, c
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
                                                          DataType::U8, DataType::S8, DataType::QASYMM8,
                                                          DataType::U16, DataType::S16, DataType::F16,
-                                                         DataType::U32, DataType::F32);
+                                                         DataType::U32, DataType::S32, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) + width_offset > output->dimension(0));
 
diff --git a/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
new file mode 100644
index 0000000000..05e05a5e57
--- /dev/null
+++ b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
@@ -0,0 +1,376 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+#include "arm_compute/runtime/NEON/functions/NELSTMLayerQuantized.h"
+
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
+
+#include <cmath>
+#include <memory>
+#include <tuple>
+
+namespace arm_compute
+{
+namespace
+{
+// Quantization info structures used in the LSTMQuantize layer
+const QuantizationInfo qasymm(1.f / 128.f, 128);
+const QuantizationInfo qsymm_3(8.f / 32768.f, 0);  // qsymm16 with 3 integer bit
+const QuantizationInfo qsymm_4(16.f / 32768.f, 0); // qsymm16 with 4 integer bit
+const QuantizationInfo qsymm_0(1.f / 32768.f, 0);  // qsymm16 with 0 integer bit
+} // namespace
+
+NELSTMLayerQuantized::NELSTMLayerQuantized(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _gemmlowp(), _output_stage(), _transpose_weights(), _concat_input_weights(), _concat_recurrent_weights(), _concat_weights(), _concat_inputs(),
+      _concat_bias(), _sigmoid_forget_gate(), _sigmoid_input_gate(), _sigmoid_output_gate(), _tanh_modulation_gate(), _tanh_output_state(), _add1(), _add2(), _mul1(), _mul2(), _mul3(),
+      _slice_input_tensor(), _slice_forget_tensor(), _slice_cell_tensor(), _slice_output_tensor(), _dequantize(), _quantize(), _input_to_input_weights(nullptr), _input_to_forget_weights(nullptr),
+      _input_to_cell_weights(nullptr), _input_to_output_weights(nullptr), _recurrent_to_input_weights(nullptr), _recurrent_to_forget_weights(nullptr), _recurrent_to_cell_weights(nullptr),
+      _recurrent_to_output_weights(nullptr), _input_gate_bias(nullptr), _forget_gate_bias(nullptr), _cell_bias(nullptr), _output_gate_bias(nullptr), _recurrent_weights(), _input_weights(), _weights(),
+      _input(), _weights_transposed(), _output_highp(), _output_lowp(), _bias(), _forget_gate_input(), _input_gate_input(), _output_gate_input(), _input_modulation_gate_input(), _forget_gate_output(),
+      _input_gate_output(), _output_gate_output(), _input_modulation_gate_output(), _cell_state1(), _cell_state2(), _output_state_tmp(), _output_state_out_symm(), _output_state_out_f32(),
+      _is_prepared(false)
+{
+}
+
+void NELSTMLayerQuantized::configure(const ITensor *input,
+                                     const ITensor *input_to_input_weights, const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights,
+                                     const ITensor *recurrent_to_input_weights, const ITensor *recurrent_to_forget_weights, const ITensor *recurrent_to_cell_weights, const ITensor *recurrent_to_output_weights,
+                                     const ITensor *input_gate_bias, const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias,
+                                     ITensor *cell_state_in, const ITensor *output_state_in,
+                                     ITensor *cell_state_out, ITensor *output_state_out)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
+                                 recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
+                                 input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in, cell_state_out, output_state_out);
+
+    ARM_COMPUTE_ERROR_THROW_ON(NELSTMLayerQuantized::validate(input->info(), input_to_input_weights->info(), input_to_forget_weights->info(), input_to_cell_weights->info(),
+                                                              input_to_output_weights->info(),
+                                                              recurrent_to_input_weights->info(), recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(),
+                                                              input_gate_bias->info(), forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(), cell_state_in->info(), output_state_in->info(), cell_state_out->info(), output_state_out->info()));
+
+    const int input_size  = input->info()->dimension(0);
+    const int batch_size  = input->info()->dimension(1);
+    const int output_size = input_to_input_weights->info()->dimension(1);
+
+    const QuantizationInfo qweights = input_to_input_weights->info()->quantization_info(); // Weights quantization
+
+    auto_init_if_empty(*cell_state_out->info(), TensorInfo(TensorShape(batch_size, output_size), 1, DataType::QSYMM16, qsymm_4));
+    auto_init_if_empty(*output_state_out->info(), TensorInfo(TensorShape(batch_size, output_size), 1, DataType::QASYMM8, qasymm));
+
+    _input_to_input_weights      = input_to_input_weights;
+    _input_to_forget_weights     = input_to_forget_weights;
+    _input_to_cell_weights       = input_to_cell_weights;
+    _input_to_output_weights     = input_to_output_weights;
+    _recurrent_to_input_weights  = recurrent_to_input_weights;
+    _recurrent_to_forget_weights = recurrent_to_forget_weights;
+    _recurrent_to_cell_weights   = recurrent_to_cell_weights;
+    _recurrent_to_output_weights = recurrent_to_output_weights;
+    _input_gate_bias             = input_gate_bias;
+    _forget_gate_bias            = forget_gate_bias;
+    _cell_bias                   = cell_bias;
+    _output_gate_bias            = output_gate_bias;
+
+    // Weights concatenation
+    std::vector<const ITensor *> inputs_weights_vector{ input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights };
+    std::vector<const ITensor *> recurrent_weights_vector{ recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights };
+
+    _input_weights.allocator()->init(TensorInfo(TensorShape(input_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
+    _concat_input_weights.configure(inputs_weights_vector, &_input_weights, Window::DimY);
+
+    _recurrent_weights.allocator()->init(TensorInfo(TensorShape(output_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
+    _concat_recurrent_weights.configure(recurrent_weights_vector, &_recurrent_weights, Window::DimY);
+
+    std::vector<const ITensor *> weights_vector{ &_recurrent_weights, &_input_weights };
+    _weights.allocator()->init(TensorInfo(TensorShape(output_size + input_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
+    _concat_weights.configure(weights_vector, &_weights, Window::DimX);
+    _transpose_weights.configure(&_weights, &_weights_transposed);
+
+    // Input concatenation
+    std::vector<const ITensor *> input_vector{ input, output_state_in };
+    _memory_group.manage(&_input);
+    _input.allocator()->init(TensorInfo(TensorShape(output_size + input_size, batch_size), 1, DataType::QASYMM8, qasymm));
+    _concat_inputs.configure(input_vector, &_input, Window::DimX);
+
+    // Bias concatenation
+    std::vector<const ITensor *> bias_vector{ input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias };
+    _bias.allocator()->init(TensorInfo(TensorShape(4 * output_size), 1, DataType::S32));
+    _concat_bias.configure(bias_vector, &_bias, Window::DimX);
+
+    // Invert the offset for gemmlowp
+    _input.info()->set_quantization_info(QuantizationInfo(qasymm.uniform().scale, -qasymm.uniform().offset));
+    _weights_transposed.info()->set_quantization_info(QuantizationInfo(qweights.uniform().scale, -qweights.uniform().offset));
+
+    // Run gemmlowp
+    _memory_group.manage(&_output_highp);
+    _output_highp.allocator()->init(TensorInfo(TensorShape(4 * output_size, batch_size), 1, DataType::S32));
+    _gemmlowp.configure(&_input, &_weights_transposed, nullptr, &_output_highp);
+    _input.allocator()->allocate();
+
+    // Set the offset back
+    _input.info()->set_quantization_info(QuantizationInfo(qasymm.uniform().scale, qasymm.uniform().offset));
+    _weights_transposed.info()->set_quantization_info(QuantizationInfo(qweights.uniform().scale, qweights.uniform().offset));
+
+    // multiplier = (input_scale * weights_scale) / output_scale (2 ^ (-12))
+    _output_lowp.allocator()->init(TensorInfo(_output_highp.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_3));
+
+    const float multiplier        = 4096.f * qasymm.uniform().scale * qweights.uniform().scale;
+    int         output_multiplier = 0;
+    int         output_shift      = 0;
+
+    quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+    _memory_group.manage(&_output_lowp);
+    _output_stage.configure(&_output_highp, &_bias, &_output_lowp, output_multiplier, output_shift);
+    _output_highp.allocator()->allocate();
+    _bias.allocator()->allocate();
+
+    // Get the gate tensors
+    _memory_group.manage(&_input_gate_input);
+    _slice_input_tensor.configure(&_output_lowp, &_input_gate_input, { 0, 0 }, { output_size, batch_size });
+    _memory_group.manage(&_forget_gate_input);
+    _slice_forget_tensor.configure(&_output_lowp, &_forget_gate_input, { output_size, 0 }, { 2 * output_size, batch_size });
+    _memory_group.manage(&_input_modulation_gate_input);
+    _slice_cell_tensor.configure(&_output_lowp, &_input_modulation_gate_input, { 2 * output_size, 0 }, { 3 * output_size, batch_size });
+    _memory_group.manage(&_output_gate_input);
+    _slice_output_tensor.configure(&_output_lowp, &_output_gate_input, { 3 * output_size, 0 }, { 4 * output_size, batch_size });
+    _output_lowp.allocator()->allocate();
+
+    // Forget gate
+    _memory_group.manage(&_forget_gate_output);
+    _forget_gate_output.allocator()->init(TensorInfo(_forget_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
+    _sigmoid_forget_gate.configure(&_forget_gate_input, &_forget_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
+    _forget_gate_input.allocator()->allocate();
+
+    // Input gate
+    _memory_group.manage(&_input_gate_output);
+    _input_gate_output.allocator()->init(TensorInfo(_input_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
+    _sigmoid_input_gate.configure(&_input_gate_input, &_input_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
+    _input_gate_input.allocator()->allocate();
+
+    // Input modulation gate equation
+    _memory_group.manage(&_input_modulation_gate_output);
+    _input_modulation_gate_output.allocator()->init(TensorInfo(_input_modulation_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
+    _tanh_modulation_gate.configure(&_input_modulation_gate_input, &_input_modulation_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f));
+    _input_modulation_gate_input.allocator()->allocate();
+
+    // Output gate
+    _memory_group.manage(&_output_gate_output);
+    _output_gate_output.allocator()->init(TensorInfo(_output_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
+    _sigmoid_output_gate.configure(&_output_gate_input, &_output_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
+    _output_gate_input.allocator()->allocate();
+
+    // Long term memory
+    _memory_group.manage(&_cell_state1);
+    _cell_state1.allocator()->init(TensorInfo(_forget_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_4));
+    _mul1.configure(&_forget_gate_output, cell_state_in, &_cell_state1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
+    _forget_gate_output.allocator()->allocate();
+
+    _memory_group.manage(&_cell_state2);
+    _cell_state2.allocator()->init(TensorInfo(_input_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_4));
+    _mul2.configure(&_input_gate_output, &_input_modulation_gate_output, &_cell_state2, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
+    _input_modulation_gate_output.allocator()->allocate();
+    _input_gate_output.allocator()->allocate();
+
+    _add1.configure(&_cell_state1, &_cell_state2, cell_state_out, ConvertPolicy::SATURATE);
+    _cell_state1.allocator()->allocate();
+    _cell_state2.allocator()->allocate();
+
+    // Short term memory
+    _memory_group.manage(&_output_state_tmp);
+    _output_state_tmp.allocator()->init(TensorInfo(cell_state_out->info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
+    _tanh_output_state.configure(cell_state_out, &_output_state_tmp, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f));
+
+    _memory_group.manage(&_output_state_out_symm);
+    _output_state_out_symm.allocator()->init(TensorInfo(_output_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
+    _mul3.configure(&_output_state_tmp, &_output_gate_output, &_output_state_out_symm, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
+    _output_gate_output.allocator()->allocate();
+    _output_state_tmp.allocator()->allocate();
+
+    // Requantize the output state from QSYMM16 to QASYMM8
+    _memory_group.manage(&_output_state_out_f32);
+    _output_state_out_f32.allocator()->init(TensorInfo(_output_state_out_symm.info()->tensor_shape(), 1, DataType::F32));
+    _dequantize.configure(&_output_state_out_symm, &_output_state_out_f32);
+    _output_state_out_symm.allocator()->allocate();
+
+    _quantize.configure(&_output_state_out_f32, output_state_out);
+    _output_state_out_f32.allocator()->allocate();
+}
+
+Status NELSTMLayerQuantized::validate(const ITensorInfo *input,
+                                      const ITensorInfo *input_to_input_weights, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
+                                      const ITensorInfo *recurrent_to_input_weights, const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
+                                      const ITensorInfo *input_gate_bias, 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)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_input_weights,
+                                        recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in,
+                                        output_state_in, cell_state_out, output_state_out);
+
+    const int input_size  = input->dimension(0);
+    const int batch_size  = input->dimension(1);
+    const int output_size = input_to_input_weights->dimension(1);
+
+    // Dimensionality checks
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(input_to_input_weights->num_dimensions() > 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(input_gate_bias->num_dimensions() > 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() > 2);
+
+    TensorInfo input_weights_info(input_to_input_weights->clone()->set_tensor_shape(TensorShape(input_size, output_size)).set_data_type(DataType::QASYMM8));
+    TensorInfo recurrent_weights_info(input_to_input_weights->clone()->set_tensor_shape(TensorShape(output_size, output_size)).set_data_type(DataType::QASYMM8).set_quantization_info(qasymm));
+    TensorInfo bias_info(input_gate_bias->clone()->set_tensor_shape(TensorShape(output_size)).set_data_type(DataType::S32));
+    TensorInfo output_state_info(cell_state_in->clone()->set_tensor_shape(TensorShape(output_size, batch_size)).set_data_type(DataType::QASYMM8).set_quantization_info(qasymm));
+    TensorInfo cell_state_info(cell_state_in->clone()->set_tensor_shape(TensorShape(output_size, batch_size)).set_data_type(DataType::QSYMM16).set_quantization_info(qsymm_4));
+
+    // Shape checks
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&input_weights_info, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&recurrent_weights_info, recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&bias_info, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&cell_state_info, cell_state_in);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&output_state_info, output_state_in);
+
+    // Data type checks
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input_weights_info, input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&recurrent_weights_info, recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&bias_info, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&cell_state_info, cell_state_in);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&output_state_info, output_state_in);
+
+    // Quantization checks
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&input_weights_info, input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&recurrent_weights_info, recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&cell_state_info, cell_state_in);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&output_state_info, output_state_in);
+
+    if(cell_state_out->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&cell_state_info, cell_state_out);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&cell_state_info, cell_state_out);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&cell_state_info, cell_state_out);
+    }
+
+    if(output_state_out->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&output_state_info, output_state_out);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&output_state_info, output_state_out);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&output_state_info, output_state_out);
+    }
+
+    return Status{};
+}
+
+void NELSTMLayerQuantized::run()
+{
+    prepare();
+
+    // Acquire all the temporaries
+    MemoryGroupResourceScope scope_mg(_memory_group);
+
+    // Concat and transpose the input
+    _concat_inputs.run();
+
+    // Run gemmlowp
+    _gemmlowp.run();
+    _output_stage.run();
+
+    // Slice the results
+    _slice_input_tensor.run();
+    _slice_forget_tensor.run();
+    _slice_cell_tensor.run();
+    _slice_output_tensor.run();
+
+    // Gates
+    // Forget gate
+    _sigmoid_forget_gate.run();
+
+    // Input gate
+    _sigmoid_input_gate.run();
+
+    // Input modulation gate
+    _tanh_modulation_gate.run();
+
+    // Output gate
+    _sigmoid_output_gate.run();
+
+    // Cell state (long term memory)
+    _mul1.run();
+    _mul2.run();
+    _add1.run();
+
+    // Output state (short term memory)
+    _tanh_output_state.run();
+    _mul3.run();
+
+    // Requantize output state from QSYMM16 to QASYMM16
+    _dequantize.run();
+    _quantize.run();
+}
+
+void NELSTMLayerQuantized::prepare()
+{
+    if(!_is_prepared)
+    {
+        _input_weights.allocator()->allocate();
+        _concat_input_weights.run();
+
+        _input_to_input_weights->mark_as_unused();
+        _input_to_forget_weights->mark_as_unused();
+        _input_to_cell_weights->mark_as_unused();
+        _input_to_output_weights->mark_as_unused();
+
+        _recurrent_weights.allocator()->allocate();
+        _concat_recurrent_weights.run();
+        _recurrent_to_input_weights->mark_as_unused();
+        _recurrent_to_forget_weights->mark_as_unused();
+        _recurrent_to_cell_weights->mark_as_unused();
+        _recurrent_to_output_weights->mark_as_unused();
+
+        _weights.allocator()->allocate();
+        _concat_weights.run();
+
+        _input_weights.mark_as_unused();
+        _input_weights.allocator()->free();
+        _recurrent_weights.mark_as_unused();
+        _recurrent_weights.allocator()->free();
+
+        _weights_transposed.allocator()->allocate();
+        _transpose_weights.run();
+
+        _weights.mark_as_unused();
+        _weights.allocator()->free();
+
+        _bias.allocator()->allocate();
+        _concat_bias.run();
+        _input_gate_bias->mark_as_unused();
+        _forget_gate_bias->mark_as_unused();
+        _cell_bias->mark_as_unused();
+        _output_gate_bias->mark_as_unused();
+
+        _is_prepared = true;
+    }
+}
+
+} // namespace arm_compute