Port NEGEMMLowp Part 1

Details:
Port NEGEMMLowpQuantizeDownInt32ScaleKernel to CpuGemmLowpQuantizeDownInt32ScaleKernel
Port NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel to CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
Port NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel to CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
Port NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel to CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
Port NEGEMMLowpOutputStage functions to CpuGemmLowpOutputStage operators

Partially Resolves: COMPMID-4403

Change-Id: I6d5f45e43f35d731d564ed3b5c0e804d2a318fb1
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5833
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

8 files changed, 1468 insertions, 0 deletions

diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp
new file mode 100644
index 0000000000..f1c797244a
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp
@@ -0,0 +1,326 @@
+/*
+ * Copyright (c) 2020-2021 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 "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
+#include "src/core/AccessWindowStatic.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_max_bound > std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type)));
+    ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_min_bound < std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))
+                                || output_stage->gemmlowp_min_bound > output_stage->gemmlowp_max_bound);
+
+    // Check biases if exist
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
+    }
+
+    if(dst->total_size() != 0)
+    {
+        if(dst->data_type() != output_stage->output_data_type && (output_stage->output_data_type == DataType::QASYMM8 || output_stage->output_data_type == DataType::QASYMM8_SIGNED))
+        {
+            ARM_COMPUTE_RETURN_ERROR_MSG("Mismatching data types");
+        }
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
+    }
+
+    return Status{};
+}
+
+inline void scale_input(int32x4x4_t &in_s32, int32x4_t result_offset_s32, int32_t result_mult_int)
+{
+    // Add the offset terms to GEMM's result
+    in_s32.val[0] = vaddq_s32(in_s32.val[0], result_offset_s32);
+    in_s32.val[1] = vaddq_s32(in_s32.val[1], result_offset_s32);
+    in_s32.val[2] = vaddq_s32(in_s32.val[2], result_offset_s32);
+    in_s32.val[3] = vaddq_s32(in_s32.val[3], result_offset_s32);
+
+    // Multiply by result_mult_int
+    in_s32.val[0] = vmulq_n_s32(in_s32.val[0], result_mult_int);
+    in_s32.val[1] = vmulq_n_s32(in_s32.val[1], result_mult_int);
+    in_s32.val[2] = vmulq_n_s32(in_s32.val[2], result_mult_int);
+    in_s32.val[3] = vmulq_n_s32(in_s32.val[3], result_mult_int);
+}
+
+template <typename T>
+inline typename std::enable_if<std::is_same<T, uint8_t>::value,
+       typename wrapper::traits::neon_vector<T, 16>::type>::type
+       convert_to_8bit(const int16x8x2_t in_s16)
+{
+    return wrapper::vcombine(wrapper::vqmovun(in_s16.val[0]), wrapper::vqmovun(in_s16.val[1]));
+}
+
+template <typename T>
+inline typename std::enable_if<std::is_same<T, int8_t>::value,
+       typename wrapper::traits::neon_vector<T, 16>::type>::type
+       convert_to_8bit(const int16x8x2_t in_s16)
+{
+    return wrapper::vcombine(wrapper::vqmovn(in_s16.val[0]), wrapper::vqmovn(in_s16.val[1]));
+}
+
+template <typename T>
+inline typename wrapper::traits::neon_vector<T, 16>::type finalize_quantization(int32x4x4_t &in_s32, int32x4_t result_shift_s32, typename wrapper::traits::neon_vector<T, 16>::type min,
+                                                                                typename wrapper::traits::neon_vector<T, 16>::type max)
+{
+    // Shift final result (negative value shift right)
+    in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32);
+    in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32);
+    in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32);
+    in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32);
+
+    // Convert S32 to S16
+    const int16x8x2_t in_s16 =
+    {
+        {
+            vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
+            vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))
+        }
+    };
+
+    // Convert S16 to S8 or U8
+    typename wrapper::traits::neon_vector<T, 16>::type out = convert_to_8bit<T>(in_s16);
+
+    out = wrapper::vmax(out, min);
+    out = wrapper::vmin(out, max);
+
+    return out;
+}
+} // namespace
+
+template <typename T>
+void CpuGemmLowpQuantizeDownInt32ScaleKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
+{
+    using VectorType = typename wrapper::traits::neon_vector<T, 16>::type;
+
+    const int32x4_t result_offset_s32 = vdupq_n_s32(_output_stage->gemmlowp_offset);
+    const int32x4_t result_shift_s32  = vdupq_n_s32(-_output_stage->gemmlowp_shift);
+    const int       window_step_x     = 16;
+    const auto      window_start_x    = static_cast<int>(window.x().start());
+    const auto      window_end_x      = static_cast<int>(window.x().end());
+
+    const int clamp_min = (_is_bounded_relu) ? _output_stage->gemmlowp_min_bound : std::numeric_limits<T>::lowest();
+    const int clamp_max = (_is_bounded_relu) ? _output_stage->gemmlowp_max_bound : std::numeric_limits<T>::max();
+
+    VectorType min = wrapper::vdup_n(static_cast<T>(clamp_min), wrapper::traits::vector_128_tag{});
+    VectorType max = wrapper::vdup_n(static_cast<T>(clamp_max), wrapper::traits::vector_128_tag{});
+
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator in(src, win);
+    Iterator out(dst, win);
+
+    if(bias != nullptr)
+    {
+        Window win_biases;
+        win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
+        win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+        Iterator bias_i(bias, win_biases);
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            // Compute 16 elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                int32x4x4_t in_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
+                    }
+                };
+
+                const int32x4x4_t bias_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12)
+                    }
+                };
+
+                // Add the bias to GEMM's result
+                in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
+                in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
+                in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]);
+                in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
+
+                // Add the offset terms to GEMM's result and multiply by result_mult_int
+                scale_input(in_s32, result_offset_s32, _output_stage->gemmlowp_multiplier);
+
+                wrapper::vstore(reinterpret_cast<T *>(out.ptr() + x), finalize_quantization<T>(in_s32, result_shift_s32, min, max));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const int bias_value = *(reinterpret_cast<const int *>(bias_i.ptr()) + x);
+                int       in_value   = *(reinterpret_cast<const int *>(in.ptr()) + x);
+
+                // Quantize
+                in_value = ((in_value + bias_value + _output_stage->gemmlowp_offset) * _output_stage->gemmlowp_multiplier) >> _output_stage->gemmlowp_shift;
+
+                // Store the result
+                *(out.ptr() + x) = static_cast<T>(utility::clamp<int>(in_value, clamp_min, clamp_max));
+            }
+        },
+        in, bias_i, out);
+    }
+    else
+    {
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            // Compute 16 elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                int32x4x4_t in_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
+                    }
+                };
+
+                // Add the offset terms to GEMM's result and multiply by result_mult_int
+                scale_input(in_s32, result_offset_s32, _output_stage->gemmlowp_multiplier);
+
+                wrapper::vstore(reinterpret_cast<T *>(out.ptr() + x), finalize_quantization<T>(in_s32, result_shift_s32, min, max));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                int in_value = *(reinterpret_cast<const int *>(in.ptr()) + x);
+
+                // Quantize
+                in_value = ((in_value + _output_stage->gemmlowp_offset) * _output_stage->gemmlowp_multiplier) >> _output_stage->gemmlowp_shift;
+
+                // Store the result
+                *(out.ptr() + x) = static_cast<T>(utility::clamp<int>(in_value, clamp_min, clamp_max));
+            }
+        },
+        in, out);
+    }
+}
+
+void CpuGemmLowpQuantizeDownInt32ScaleKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage)
+{
+    ARM_COMPUTE_UNUSED(bias);
+    // Perform validate step
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst, output_stage);
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*dst, src->clone()->set_data_type(output_stage->output_data_type));
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src,
+                                                  bias,
+                                                  dst,
+                                                  output_stage));
+
+    _output_stage = output_stage;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*src, Steps());
+
+    ICpuKernel::configure(win);
+
+    // Check if we need to clamp the result using min and max
+    _is_bounded_relu = ((_output_stage->gemmlowp_min_bound != _output_stage->gemmlowp_max_bound)
+                        && !(_output_stage->gemmlowp_min_bound == std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))
+                             && _output_stage->gemmlowp_max_bound == std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))));
+    if(_output_stage->output_data_type == DataType::QASYMM8)
+    {
+        _func = &CpuGemmLowpQuantizeDownInt32ScaleKernel::run_internal<uint8_t>;
+    }
+    else if(_output_stage->output_data_type == DataType::QASYMM8_SIGNED)
+    {
+        _func = &CpuGemmLowpQuantizeDownInt32ScaleKernel::run_internal<int8_t>;
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Data type not supported");
+    }
+}
+
+Status CpuGemmLowpQuantizeDownInt32ScaleKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, output_stage));
+    return Status{};
+}
+
+void CpuGemmLowpQuantizeDownInt32ScaleKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
+
+    auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
+    auto dst  = tensors.get_tensor(TensorType::ACL_DST);
+    (this->*_func)(src, bias, dst, window);
+}
+
+const char *CpuGemmLowpQuantizeDownInt32ScaleKernel::name() const
+{
+    return "CpuGemmLowpQuantizeDownInt32ScaleKernel";
+}
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h
new file mode 100644
index 0000000000..646242dc7e
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h
@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2020-2021 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_CPU_GEMMLOWP_QUANTIZEDOWNINT32_SCALE_KERNEL_H
+#define ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32_SCALE_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+namespace cpu
+{
+namespace kernels
+{
+/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8/QASYMM8_SIGNED
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Add offset terms to final result
+ *  -# Multiply each entry of result by result_mult_int
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Shift the int32 accumulator by result_shift
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values:
+ *  -#  -to the [0..255] range and cast to QASYMM8.
+ *  -#  -to the [-128..127] range and cast to QASYMM8_SIGNED.
+ *
+ */
+class CpuGemmLowpQuantizeDownInt32ScaleKernel : public ICpuKernel
+{
+public:
+    /** Default constructor */
+    CpuGemmLowpQuantizeDownInt32ScaleKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpQuantizeDownInt32ScaleKernel);
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  src          Input tensor info. Data type supported: S32
+     * @param[in]  bias         Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] dst          Output tensor info. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[out] output_stage GEMMLowp output stage metadata.
+     */
+    void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to CpuGemmLowpQuantizeDownInt32ScaleKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    /** Template function to run the NEGEMMLowpQuantizeDownInt32ScaleKernel
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Biases tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel. (Must be a valid region of the window returned by window())
+     */
+    template <typename T>
+    void run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    /** Common signature for all the specialised CpuGemmLowpQuantizeDownInt32ScaleKernel functions
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Biases tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (CpuGemmLowpQuantizeDownInt32ScaleKernel::*)(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    QuantizeDownFunctionPtr        _func{ nullptr };
+    const GEMMLowpOutputStageInfo *_output_stage{ nullptr };
+    bool                           _is_bounded_relu{ false };
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32_SCALE_KERNEL_H */
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
new file mode 100644
index 0000000000..390e269cbb
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
@@ -0,0 +1,227 @@
+/*
+ * Copyright (c) 2019-2021 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 "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "src/core/NEON/NESymm.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
+
+    // Check biases if exist
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
+    }
+
+    if(dst->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QSYMM16);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, src);
+    }
+
+    return Status{};
+}
+} // namespace
+
+template <bool is_bounded_relu>
+void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
+{
+    const int16x8_t min_s16 = vdupq_n_s16(static_cast<int16_t>(_min));
+    const int16x8_t max_s16 = vdupq_n_s16(static_cast<int16_t>(_max));
+
+    ARM_COMPUTE_UNUSED(min_s16);
+    ARM_COMPUTE_UNUSED(max_s16);
+
+    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());
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator in(src, win_collapsed);
+    Iterator out(dst, win_collapsed);
+    if(bias != nullptr)
+    {
+        Window win_biases;
+        win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
+        win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+        Iterator bias_i(bias, win_biases);
+        execute_window_loop(win_collapsed, [&](const Coordinates &)
+        {
+            // Compute 16 elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                int32x4x2_t in_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
+                    }
+                };
+
+                const int32x4x2_t bias_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4)
+                    }
+                };
+
+                // Add the bias to GEMM's result
+                in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
+                in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
+
+                vst1q_s16(reinterpret_cast<int16_t *>(out.ptr()) + x, finalize_quantization_int16<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, min_s16, max_s16));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x);
+                int32_t       in_value   = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
+
+                // Add bias
+                in_value += bias_value;
+                // Finalize and store the result
+                *(reinterpret_cast<int16_t *>(out.ptr()) + x) = finalize_quantization_int16<is_bounded_relu>(in_value, _result_fixedpoint_multiplier, _result_shift, static_cast<int16_t>(_min),
+                                                                                                             static_cast<int16_t>(_max));
+            }
+        },
+        in, out, bias_i);
+    }
+    else
+    {
+        execute_window_loop(win_collapsed, [&](const Coordinates &)
+        {
+            // Compute 16 elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                int32x4x2_t in_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
+                    }
+                };
+
+                vst1q_s16(reinterpret_cast<int16_t *>(out.ptr()) + x, finalize_quantization_int16<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, min_s16, max_s16));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
+                ARM_COMPUTE_UNUSED(in_value);
+                // Finalize and store the result
+                *(reinterpret_cast<int16_t *>(out.ptr()) + x) = finalize_quantization_int16<is_bounded_relu>(in_value, _result_fixedpoint_multiplier, _result_shift, static_cast<int16_t>(_min),
+                                                                                                             static_cast<int16_t>(_max));
+            }
+        },
+        in, out);
+    }
+}
+
+void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift,
+                                                                           int min, int max)
+{
+    // Perform validate step
+    ARM_COMPUTE_UNUSED(bias, dst);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, min, max));
+
+    _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
+    _result_shift                 = result_shift;
+    _min                          = min;
+    _max                          = max;
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*src, src->clone()->set_data_type(DataType::QSYMM16));
+    // Configure kernel window
+    Window win_config = calculate_max_window(*src, Steps());
+    ICpuKernel::configure(win_config);
+
+    // Check if we need to clamp the result using min and max
+    const bool is_bounded_relu = !(min <= -32768 && max >= 32767);
+    _func                      = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<true> :
+                                 &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<false>;
+}
+
+Status CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
+    return Status{};
+}
+
+void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
+
+    auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
+    auto dst  = tensors.get_tensor(TensorType::ACL_DST);
+
+    (this->*_func)(src, bias, dst, window);
+}
+
+const char *CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::name() const
+{
+    return "CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel";
+}
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..4d743e9a0a
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2019-2021 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_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT16_SCALEBYFIXEDPOINT_KERNEL_H
+#define ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT16_SCALEBYFIXEDPOINT_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+namespace cpu
+{
+namespace kernels
+{
+/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QSYMM16
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QSYMM16 value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Round to nearest division by a power-of-two using result_shift
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [-32768, 32767] range and cast to QSYMM16.
+ *
+ */
+class CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel : public ICpuKernel
+{
+public:
+    /** Default constructor */
+    CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel);
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  src                          Input tensor info. Data type supported: S32
+     * @param[in]  bias                         Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] dst                          Output tensor info. Data type supported: Data type supported: QSYMM16
+     * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
+     * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
+     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to 0.
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QSYMM16.
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to 0.
+     */
+    void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min = 0, int max = 0);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    /** Template function to run the CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Bias tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <bool is_bounded_relu>
+    void run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    /** Common signature for all the specialised CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel functions
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Bias tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::*)(
+                                        const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    QuantizeDownFunctionPtr _func{ nullptr };
+    int                     _result_fixedpoint_multiplier{ 0 };
+    int                     _result_shift{ 0 };
+    int                     _min{ 0 };
+    int                     _max{ 0 };
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT16_SCALEBYFIXEDPOINT_KERNEL_H */
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
new file mode 100644
index 0000000000..318b6a06f8
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
@@ -0,0 +1,239 @@
+/*
+ * Copyright (c) 2019-2021 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 "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "src/core/NEON/NEAsymm.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
+
+    // Check biases if exist
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
+    }
+
+    if(dst->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8_SIGNED);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, src);
+    }
+
+    return Status{};
+}
+} // namespace
+
+template <bool is_bounded_relu>
+void CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
+{
+    const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift);
+    const int8x16_t min_s8                        = vdupq_n_s8(static_cast<int8_t>(_min));
+    const int8x16_t max_s8                        = vdupq_n_s8(static_cast<int8_t>(_max));
+
+    ARM_COMPUTE_UNUSED(min_s8, max_s8);
+
+    const int  window_step_x  = 16;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator in(src, win_collapsed);
+    Iterator out(dst, win_collapsed);
+    if(bias != nullptr)
+    {
+        Window win_biases;
+        win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
+        win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+        Iterator bias_i(bias, win_biases);
+        execute_window_loop(win_collapsed, [&](const Coordinates &)
+        {
+            // Compute 16 elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                int32x4x4_t in_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
+                    }
+                };
+
+                const int32x4x4_t bias_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12)
+                    }
+                };
+
+                // Add the bias to GEMM's result
+                in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
+                in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
+                in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]);
+                in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
+
+                vst1q_s8(reinterpret_cast<int8_t *>(out.ptr() + x),
+                         finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_s8, max_s8, is_bounded_relu));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x);
+                int32_t       in_value   = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
+
+                // Add bias
+                in_value += bias_value;
+                // Finalize and store the result
+                *reinterpret_cast<int8_t *>(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift,
+                                                                                   static_cast<int8_t>(_min), static_cast<int8_t>(_max), is_bounded_relu);
+            }
+        },
+        in, out, bias_i);
+    }
+    else
+    {
+        execute_window_loop(win_collapsed, [&](const Coordinates &)
+        {
+            // Compute 16 elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                int32x4x4_t in_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
+                    }
+                };
+
+                vst1q_s8(reinterpret_cast<int8_t *>(out.ptr() + x),
+                         finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_s8, max_s8, is_bounded_relu));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
+
+                // Finalize and store the result
+                *reinterpret_cast<int8_t *>(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift,
+                                                                                   static_cast<int8_t>(_min), static_cast<int8_t>(_max), is_bounded_relu);
+            }
+        },
+        in, out);
+    }
+}
+
+void CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift,
+                                                                          int result_offset_after_shift, int min, int max)
+{
+    ARM_COMPUTE_UNUSED(bias);
+    // Perform validate step
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, min, max));
+
+    _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
+    _result_shift                 = result_shift;
+    _result_offset_after_shift    = result_offset_after_shift;
+    _min                          = min;
+    _max                          = max;
+
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*dst, src->clone()->set_data_type(DataType::QASYMM8_SIGNED));
+
+    // Configure kernel window
+    Window win_config = calculate_max_window(*src, Steps());
+    ICpuKernel::configure(win_config);
+
+    // Check if we need to clamp the result using min and max
+    const bool is_bounded_relu = !(min <= -128 && max >= 127);
+    _func                      = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run_internal<true> :
+                                 &CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run_internal<false>;
+}
+
+Status CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, min, max));
+    return Status{};
+}
+
+void CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
+
+    auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
+    auto dst  = tensors.get_tensor(TensorType::ACL_DST);
+
+    (this->*_func)(src, bias, dst, window);
+}
+
+const char *CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::name() const
+{
+    return "CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel";
+}
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..a941f1f542
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
@@ -0,0 +1,114 @@
+/*
+ * Copyright (c) 2019-2021 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_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT8_SCALEBYFIXEDPOINT_KERNEL_H
+#define ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT8_SCALEBYFIXEDPOINT_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+namespace cpu
+{
+namespace kernels
+{
+/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8_SIGNED
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8_SIGNED value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Round to nearest division by a power-of-two using result_shift
+ *  -# Add offset to each result
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [-128..127] range and cast to QASYMM8_SIGNED.
+ *
+ */
+class CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel : public ICpuKernel
+{
+public:
+    /** Default constructor */
+    CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel);
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  src                          Input tensor info. Data type supported: S32
+     * @param[in]  bias                         Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] dst                          Output tensor info. Data type supported: Data type supported: QASYMM8_SIGNED
+     * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
+     * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
+     * @param[in]  result_offset_after_shift    Offset to be applied to result before converting it back to QASYMM8_SIGNED
+     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     */
+    void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min = 0, int max = 0);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    /** Template function to run the CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Bias tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <bool is_bounded_relu>
+    void run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    /** Common signature for all the specialised CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel functions
+     *
+     * @param[in]  src    Input tensor info
+     * @param[in]  bias   Bias tensor info
+     * @param[out] dst    Output tensor info
+     * @param[in]  window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::*)(
+                                        const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    QuantizeDownFunctionPtr _func{ nullptr };
+    int                     _result_fixedpoint_multiplier{ 0 };
+    int                     _result_shift{ 0 };
+    int                     _result_offset_after_shift{ 0 };
+    int                     _min{ 0 };
+    int                     _max{ 0 };
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOINT8_SCALEBYFIXEDPOINT_KERNEL_H */
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
new file mode 100644
index 0000000000..6631a4fc67
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -0,0 +1,236 @@
+/*
+ * Copyright (c) 2017-2021 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 "src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "src/core/NEON/NEAsymm.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
+
+    // Check biases if exist
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
+    }
+
+    if(dst->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, src);
+    }
+
+    return Status{};
+}
+} // namespace
+
+template <bool is_bounded_relu>
+void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
+{
+    const int32x4_t  result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift);
+    const uint8x16_t min_u8                        = vdupq_n_u8(static_cast<uint8_t>(_min));
+    const uint8x16_t max_u8                        = vdupq_n_u8(static_cast<uint8_t>(_max));
+
+    ARM_COMPUTE_UNUSED(min_u8);
+    ARM_COMPUTE_UNUSED(max_u8);
+
+    const int  window_step_x  = 16;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator in(src, win_collapsed);
+    Iterator out(dst, win_collapsed);
+    if(bias != nullptr)
+    {
+        Window win_biases;
+        win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
+        win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+        Iterator bias_i(bias, win_biases);
+        execute_window_loop(win_collapsed, [&](const Coordinates &)
+        {
+            // Compute 16 elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                int32x4x4_t in_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
+                    }
+                };
+
+                const int32x4x4_t bias_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12)
+                    }
+                };
+
+                // Add the bias to GEMM's result
+                in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
+                in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
+                in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]);
+                in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
+
+                vst1q_u8(out.ptr() + x, finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x);
+                int32_t       in_value   = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
+
+                // Add bias
+                in_value += bias_value;
+                // Finalize and store the result
+                *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast<uint8_t>(_min), static_cast<uint8_t>(_max), is_bounded_relu);
+            }
+        },
+        in, out, bias_i);
+    }
+    else
+    {
+        execute_window_loop(win_collapsed, [&](const Coordinates &)
+        {
+            // Compute 16 elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                int32x4x4_t in_s32 =
+                {
+                    {
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
+                    }
+                };
+
+                vst1q_u8(out.ptr() + x, finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
+
+                // Finalize and store the result
+                *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast<uint8_t>(_min), static_cast<uint8_t>(_max), is_bounded_relu);
+            }
+        },
+        in, out);
+    }
+}
+
+void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift,
+                                                                           int result_offset_after_shift, int min, int max)
+{
+    ARM_COMPUTE_UNUSED(bias);
+    // Perform validate step
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, min, max));
+
+    _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
+    _result_shift                 = result_shift;
+    _result_offset_after_shift    = result_offset_after_shift;
+    _min                          = min;
+    _max                          = max;
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*dst, src->clone()->set_data_type(DataType::QASYMM8));
+
+    // Configure kernel window
+    auto win_config = calculate_max_window(*src, Steps());
+    ICpuKernel::configure(win_config);
+
+    // Check if we need to clamp the result using min and max
+    const bool is_bounded_relu = !(min <= 0 && max >= 255);
+    _func                      = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<true> :
+                                 &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<false>;
+}
+
+Status CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, min, max));
+    return Status{};
+}
+
+void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
+
+    auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
+    auto dst  = tensors.get_tensor(TensorType::ACL_DST);
+
+    (this->*_func)(src, bias, dst, window);
+}
+
+const char *CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::name() const
+{
+    return "CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel";
+}
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..9b4c056419
--- /dev/null
+++ b/src/core/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h
@@ -0,0 +1,108 @@
+/*
+ * Copyright (c) 2017-2021 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_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOUINT8_SCALEBYFIXEDPOINT_KERNEL_H
+#define ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOUINT8_SCALEBYFIXEDPOINT_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+namespace cpu
+{
+namespace kernels
+{
+/** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Round to nearest division by a power-of-two using result_shift
+ *  -# Add offset to each result
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
+ *
+ */
+class CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel : public ICpuKernel
+{
+public:
+    /** Default constructor */
+    CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel);
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  src                          Input tensor info. Data type supported: S32
+     * @param[in]  bias                         Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] dst                          Output tensor info. Data type supported: Data type supported: QASYMM8
+     * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
+     * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
+     * @param[in]  result_offset_after_shift    Offset to be applied to result before converting it back to QASYMM8
+     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QASYMM8
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     */
+    void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min = 0, int max = 0);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    /** Template function to run the CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
+     *
+     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <bool is_bounded_relu>
+    void run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    /** Common signature for all the specialised CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel functions
+     *
+     * @param[in] window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::*)(
+                                        const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
+
+    QuantizeDownFunctionPtr _func{ nullptr };
+    int                     _result_fixedpoint_multiplier{ 0 };
+    int                     _result_shift{ 0 };
+    int                     _result_offset_after_shift{ 0 };
+    int                     _min{ 0 };
+    int                     _max{ 0 };
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWNINT32TOUINT8_SCALEBYFIXEDPOINT_KERNEL_H */