COMPMID-2406: Create a new GEMMLowpQuantizeDownInt32ToInt16ScaleKernel for NEON

Change-Id: I3f3e247728fd6dafca066e41835f0ef9442d9b7a
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1379
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

10 files changed, 807 insertions, 8 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 206ff06f8a..4023d82107 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -78,6 +78,7 @@
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h"
diff --git a/arm_compute/core/NEON/NESymm.h b/arm_compute/core/NEON/NESymm.h
new file mode 100644
index 0000000000..0479753426
--- /dev/null
+++ b/arm_compute/core/NEON/NESymm.h
@@ -0,0 +1,106 @@
+/*
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_NESYMM_H__
+#define __ARM_COMPUTE_NESYMM_H__
+
+#include "NEAsymm.h"
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+/** Performs final quantization step on 8 signed 16-bit elements
+ *
+ * @tparam is_bounded_relu Specified if a fused bounded relu should be applied
+ *
+ * @param[in] in_s32                       Input to be quantized.
+ * @param[in] result_fixedpoint_multiplier Result multiplier parameter
+ * @param[in] result_shift                 Result shift parameter
+ * @param[in] min_s16                      Relu lower bound
+ * @param[in] max_s16                      Relu upper bound
+ *
+ * @return Quantized values
+ */
+template <bool is_bounded_relu>
+int16x8_t finalize_quantization_int16(int32x4x2_t &in_s32,
+                                      int          result_fixedpoint_multiplier,
+                                      int32_t      result_shift,
+                                      int16x8_t    min_s16,
+                                      int16x8_t    max_s16)
+{
+    // Fixed point multiplication with vector saturating rounding doubling multiply high with scalar
+    in_s32.val[0] = vqrdmulhq_n_s32(in_s32.val[0], result_fixedpoint_multiplier);
+    in_s32.val[1] = vqrdmulhq_n_s32(in_s32.val[1], result_fixedpoint_multiplier);
+
+    // Round to the nearest division by a power-of-two using result_shift_s32
+    in_s32.val[0] = rounding_divide_by_pow2(in_s32.val[0], result_shift);
+    in_s32.val[1] = rounding_divide_by_pow2(in_s32.val[1], result_shift);
+
+    // Convert S32 to S16
+    int16x8_t out_s16 = vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1]));
+
+    if(is_bounded_relu)
+    {
+        out_s16 = vmaxq_s16(out_s16, min_s16);
+        out_s16 = vminq_s16(out_s16, max_s16);
+    }
+
+    return out_s16;
+}
+
+/** Performs final quantization step on single signed 16-bit element
+ *
+ * @tparam is_bounded_relu Specified if a fused bounded relu should be applied
+ *
+ * @param[in] in_value                     Input to be quantized.
+ * @param[in] result_fixedpoint_multiplier Result multiplier parameter
+ * @param[in] result_shift                 Result shift parameter
+ * @param[in] min_s16                      Relu lower bound
+ * @param[in] max_s16                      Relu upper bound
+ *
+ * @return Quantized values
+ */
+template <bool is_bounded_relu>
+inline int16_t finalize_quantization_int16(int32_t in_value, int result_fixedpoint_multiplier,
+                                           int32_t result_shift, int16_t min_s16, int16_t max_s16)
+{
+    int32x4_t in_s32 = vdupq_n_s32(in_value);
+
+    // Fixed point multiplication with vector saturating rounding doubling multiply high with scalar
+    in_value = vgetq_lane_s32(vqrdmulhq_n_s32(in_s32, result_fixedpoint_multiplier), 0);
+
+    // Shift value by result_shift_s32
+    in_value = rounding_divide_by_pow2(in_value, result_shift);
+
+    // Bound the result
+    int16_t out_s16 = static_cast<int16_t>(std::max<int32_t>(-32768, std::min<int32_t>(32767, in_value)));
+
+    if(is_bounded_relu)
+    {
+        out_s16 = static_cast<int16_t>(std::max(min_s16, std::min(max_s16, out_s16)));
+    }
+
+    return out_s16;
+}
+} // namespace arm_compute
+#endif // __ARM_COMPUTE_NESYMM_H__
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..806c0e4c2f
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
@@ -0,0 +1,116 @@
+/*
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H__
+#define __ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** NEON 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 NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel";
+    }
+    /** Constructor */
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  input                        Input tensor. Data type supported: S32
+     * @param[in]  bias                         Biases tensor. 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] output                       Output tensor. 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(const ITensor *input, const ITensor *bias, ITensor *output, 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 of @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+     *
+     * @param[in] input  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 info with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[in] output Output tensor info. Data type supported: Data type supported: QSYMM16
+     * @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.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    /** Template function to run the NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+     *
+     * @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(const Window &window);
+
+    /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel functions
+     *
+     * @param[in] window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::*)(const Window &window);
+
+    QuantizeDownFunctionPtr _func;
+    const ITensor          *_input;
+    const ITensor          *_bias;
+    ITensor                *_output;
+    int                     _result_fixedpoint_multiplier;
+    int                     _result_shift;
+    int                     _min;
+    int                     _max;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
index 77bfb98ba6..5ece753660 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -147,5 +147,63 @@ public:
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
 };
+/** Basic function to execute NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint on NEON.
+ *
+ *  NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint depends on 2 parameters:
+ *
+ *  result_fixedpoint_multiplier, result_shift
+ *
+ * The final result is:
+ *
+ * (FixedPointMul(input[i][k], result_fixedpoint_multiplier) >> result_shift)
+ *
+ * where FixedPointMul(x, y) is the nearest integer to the following
+ * mathematical expression, evaluated without overflow or intermediate rounding:
+ *
+ * (x * y) / 2^31
+ *
+ * For more information: https://github.com/google/gemmlowp/blob/master/public/output_stages.h#L68
+ *
+ * In case the bias tensor is provided, the final result is:
+ *
+ * ((FixedPointMul(input[i][k] + bias[k], result_fixedpoint_multiplier)) >> result_shift) + result_offset_after_shift
+ *
+ *  This function calls the following NEON kernels:
+ *
+ * -# @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+ *
+ * @note The function accepts also 2 optional input arguments (min and max) which can be used to implement "rectified linear unit" activation functions
+ *       after the result is shifted right by result_shift
+*/
+class NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint : public INESimpleFunctionNoBorder
+{
+public:
+    /** Initialise the kernel's inputs, output
+     *
+     * @param[in]  input                        Input tensor. Data type supported: S32
+     * @param[in]  bias                         Biases tensor. 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] output                       Output tensor. 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                 Number of bits to shift right 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(const ITensor *input, const ITensor *bias, ITensor *output, 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 of @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
+     *
+     * @param[in] input  Input tensor info. It is the output of @ref NEGEMMLowpMatrixMultiplyCore function. Data type supported: S32
+     * @param[in] bias   Biases tensor info. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
+     *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[in] output Output tensor info. Data type supported: Data type supported: QSYMM16
+     * @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.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
+};
 } // namespace arm_compute
-#endif /*__ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__ */
\ No newline at end of file
+#endif /*__ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__ */
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
new file mode 100644
index 0000000000..bc513e6618
--- /dev/null
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
@@ -0,0 +1,237 @@
+/*
+ * 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/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/NESymm.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 <arm_neon.h>
+#include <cstddef>
+#include <cstdint>
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
+    ARM_COMPUTE_RETURN_ERROR_ON(max > 32767);
+    ARM_COMPUTE_RETURN_ERROR_ON(min < -32768 || min > max);
+
+    // Check biases if exist
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+    }
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QSYMM16);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, input);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_data_type(DataType::QSYMM16));
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input, Steps());
+
+    // NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(output->num_dimensions());
+    output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
+
+    return std::make_pair(Status{}, win);
+}
+} // namespace
+
+class Coordinates;
+
+template <bool is_bounded_relu>
+void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(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(_input, win_collapsed);
+    Iterator out(_output, 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(_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.ptr()) + x + 0),
+                        vld1q_s32(reinterpret_cast<const int32_t *>(bias.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.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);
+    }
+    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);
+    }
+}
+
+NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel()
+    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _min(0), _max(0)
+{
+}
+
+void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
+                                                                          int min, int max)
+{
+    // Perform validate step
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
+
+    _input                        = input;
+    _bias                         = bias;
+    _output                       = output;
+    _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
+    _result_shift                 = result_shift;
+    _min                          = min;
+    _max                          = max;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    INEKernel::configure(win_config.second);
+
+    // Check if we need to clamp the result using min and max
+    const bool is_bounded_relu = ((min != max) && !(min == -32768 && max == 32767));
+    _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<false>;
+}
+
+Status NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::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));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
+
+    return Status{};
+}
+
+void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    (this->*_func)(window);
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp b/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp
index ce69fa0bfd..b89e7a168e 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,12 +24,13 @@
 #include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
 
 #include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
 #include "support/ToolchainSupport.h"
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 void NEGEMMLowpQuantizeDownInt32ToUint8Scale::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_offset, int result_mult_int, int result_shift, int min, int max)
 {
     auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel>();
@@ -53,4 +54,17 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::configure(const ITenso
 Status NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     return NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, min, max);
-}
\ No newline at end of file
+}
+
+void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int min, int max)
+{
+    auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel>();
+    k->configure(input, bias, output, result_fixedpoint_multiplier, result_shift, min, max);
+    _kernel = std::move(k);
+}
+
+Status NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+{
+    return NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(input, bias, output, min, max);
+}
+} // namespace arm_compute
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index f0460b4a23..2f604c95ea 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -294,6 +294,9 @@ const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases = framewo
 using NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture =
     GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint>;
 
+using NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture =
+    GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>;
+
 // *INDENT-OFF*
 // clang-format off
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
@@ -404,6 +407,117 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedP
 TEST_SUITE_END() // BoundedReLu
 
 TEST_SUITE_END() // QuantizeDownInt32ToUint8ScaleByFixedPoint
+
+TEST_SUITE(QuantizeDownInt32ToInt16ScaleByFixedPoint)
+
+const auto quantize_down_int32_to_int16_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
+                                                                    2)
+                                                                    * framework::dataset::make("min", 0) * framework::dataset::make("max", 0) * framework::dataset::make("addBias", { false, true });
+
+const auto quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
+                                                                         2)
+                                                                         * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true });
+
+using NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture =
+    GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>;
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
+    framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Input not a multiple of 16
+                                             TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Invalid min and max
+                                             TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type
+                                          }),
+    framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32),
+                                            TensorInfo(TensorShape(21U), 1, DataType::S32),
+                                            TensorInfo(TensorShape(20U), 1, DataType::S32),
+                                          })),
+    framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QSYMM16),
+                                            TensorInfo(TensorShape(21U, 13U), 1, DataType::QSYMM16),
+                                            TensorInfo(TensorShape(20U, 13U), 1, DataType::S32),
+                                           })),
+    framework::dataset::make("Min",{        -205,
+                                            -60000,
+                                            -180,
+                                           })),
+    framework::dataset::make("Max",{        205,
+                                            60000,
+                                            180,
+                                           })),
+    framework::dataset::make("Expected", { true, false, false })),
+    a_info, b_info, output_info, min, max, expected)
+{
+    // Lock tensors
+    Status status =  NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(&a_info.clone()->set_is_resizable(false),
+                                                                                 &b_info.clone()->set_is_resizable(false),
+                                                                                 &output_info.clone()->set_is_resizable(false),
+                                                                                 min,
+                                                                                 max);
+    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                                                                   quantize_down_int32_to_int16_scale_by_fixedpoint_cases),
+               shape, result_fixedpoint_multiplier, result_shift, min, max, add_bias)
+{
+    TensorShape shape_bias(shape[0]);
+
+    // Create tensors
+    Tensor in   = create_tensor<Tensor>(shape, DataType::S32);
+    Tensor bias = create_tensor<Tensor>(shape_bias, DataType::S32);
+    Tensor out  = create_tensor<Tensor>(shape, DataType::QSYMM16);
+
+    ARM_COMPUTE_EXPECT(in.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(out.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint output_stage;
+    output_stage.configure(&in, add_bias ? &bias : nullptr, &out, result_fixedpoint_multiplier, result_shift, min, max);
+
+    // Validate valid region input and output
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(in.info()->valid_region(), valid_region);
+    validate(out.info()->valid_region(), valid_region);
+
+    // Validate valid region bias
+    if(add_bias)
+    {
+        const ValidRegion valid_region_bias = shape_to_valid_region(shape_bias);
+        validate(bias.info()->valid_region(), valid_region_bias);
+    }
+
+    // Validate padding
+    const PaddingSize padding(0);
+    validate(in.info()->padding(), padding);
+    validate(out.info()->padding(), padding);
+
+    if(add_bias)
+    {
+        validate(bias.info()->padding(), padding);
+    }
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int16_scale_by_fixedpoint_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+TEST_SUITE(BoundedReLu)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // BoundedReLu
+
+TEST_SUITE_END() // QuantizeDownInt32ToInt16ScaleByFixedPoint
+
 TEST_SUITE_END() // OutputStage
 
 TEST_SUITE_END() // GEMMLowp
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index ad5acfc418..8385221c78 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -385,6 +385,98 @@ protected:
     SimpleTensor<uint8_t> _reference{};
 };
 
+template <typename TensorType, typename AccessorType, typename FunctionType>
+class GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t min, int32_t max, bool add_bias)
+    {
+        _target    = compute_target(shape, result_fixedpoint_multiplier, result_shift, min, max, add_bias);
+        _reference = compute_reference(shape, result_fixedpoint_multiplier, result_shift, min, max, add_bias);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_int_distribution<> distribution(-6000, 6000);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t min, int32_t max, bool add_bias)
+    {
+        TensorShape shape_bias(shape[0]);
+
+        // Create tensors
+        TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1);
+        TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1);
+        TensorType c = create_tensor<TensorType>(shape, DataType::QSYMM16, 1);
+
+        // Create and configure function
+        FunctionType output_stage;
+        output_stage.configure(&a, add_bias ? &b : nullptr, &c, result_fixedpoint_multiplier, result_shift, min, max);
+
+        ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        a.allocator()->allocate();
+        c.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensor
+        fill(AccessorType(a), 0);
+
+        if(add_bias)
+        {
+            ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+            // Allocate bias tensor
+            b.allocator()->allocate();
+
+            ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+            // Fill tensor
+            fill(AccessorType(b), 1);
+        }
+
+        // Compute GEMM function
+        output_stage.run();
+        return c;
+    }
+
+    SimpleTensor<int16_t> compute_reference(const TensorShape &shape, int32_t result_fixed_point_multiplier, int32_t result_shift, int32_t min, int32_t max,
+                                            bool add_bias)
+    {
+        // Create reference
+        TensorShape shape_bias(shape[0]);
+
+        SimpleTensor<int32_t> a{ shape, DataType::S32, 1 };
+        SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 };
+
+        // Fill reference
+        fill(a, 0);
+
+        if(add_bias)
+        {
+            // Fill bias
+            fill(b, 1);
+
+            return reference::gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint<int32_t>(a, b, result_fixed_point_multiplier, result_shift, min, max);
+        }
+        else
+        {
+            return reference::gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint<int32_t>(a, result_fixed_point_multiplier, result_shift, min, max);
+        }
+    }
+
+    TensorType            _target{};
+    SimpleTensor<int16_t> _reference{};
+};
+
 template <typename TensorType, typename AccessorType, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
 class GEMMLowpMatrixMultiplyReshapedValidationFixture : public framework::Fixture
 {
diff --git a/tests/validation/reference/GEMMLowp.cpp b/tests/validation/reference/GEMMLowp.cpp
index 9a7e409e8a..97d05327e7 100644
--- a/tests/validation/reference/GEMMLowp.cpp
+++ b/tests/validation/reference/GEMMLowp.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -95,6 +95,34 @@ void quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> *in,
         (*dst)[i] = static_cast<uint8_t>(std::max(0, std::min(255, result)));
     }
 }
+
+template <typename T>
+void quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> *in, const SimpleTensor<T> *bias, SimpleTensor<int16_t> *dst, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                      int32_t min, int32_t max)
+{
+    const int cols_in = in->shape().x();
+
+    for(int i = 0; i < in->num_elements(); ++i)
+    {
+        int32_t result = (*in)[i];
+
+        if(bias != nullptr)
+        {
+            result += (*bias)[i % cols_in];
+        }
+
+        // Fixed point multiplication
+        result = asymm_rounding_divide_by_pow2(asymm_int_mult(result, result_fixedpoint_multiplier), result_shift);
+
+        // Bounded ReLu
+        if(min != max)
+        {
+            result = std::max(min, std::min(max, result));
+        }
+
+        (*dst)[i] = static_cast<int16_t>(std::max(-32768, std::min(32767, result)));
+    }
+}
 } // namespace
 
 template <typename T_out, typename T_in>
@@ -201,10 +229,36 @@ SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(
     return dst;
 }
 
+template <typename T>
+SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t min,
+                                                                                int32_t max)
+{
+    SimpleTensor<int16_t> dst(in.shape(), DataType::QSYMM16);
+
+    quantize_down_int32_to_int16_scale_by_fixedpoint<T>(&in, nullptr, &dst, result_fixedpoint_multiplier, result_shift, min, max);
+
+    return dst;
+}
+
+template <typename T>
+SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                                                int32_t min, int32_t max)
+{
+    SimpleTensor<int16_t> dst(in.shape(), DataType::QSYMM16);
+
+    quantize_down_int32_to_int16_scale_by_fixedpoint<T>(&in, &bias, &dst, result_fixedpoint_multiplier, result_shift, min, max);
+
+    return dst;
+}
+
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, int32_t result_fixedpoint_multiplier, int32_t result_shift,
                                                                                          int32_t result_offset_after_shift, int32_t min, int32_t max);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_fixedpoint_multiplier,
                                                                                          int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
+template SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                                                         int32_t min, int32_t max);
+template SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_fixedpoint_multiplier,
+                                                                                         int32_t result_shift, int32_t min, int32_t max);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min,
                                                                            int32_t max);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_offset, int32_t result_mult_int,
diff --git a/tests/validation/reference/GEMMLowp.h b/tests/validation/reference/GEMMLowp.h
index 4396155b96..5581f67652 100644
--- a/tests/validation/reference/GEMMLowp.h
+++ b/tests/validation/reference/GEMMLowp.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -58,6 +58,13 @@ SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(
 template <typename T>
 SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_fixedpoint_multiplier, int32_t result_shift,
                                                                                 int32_t result_offset_after_shift, int32_t min = 0, int32_t max = 0);
+
+template <typename T>
+SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, int32_t result_fixedpoint_multiplier, int32_t result_shift,
+                                                                                int32_t min, int32_t max);
+template <typename T>
+SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_fixedpoint_multiplier,
+                                                                                int32_t result_shift, int32_t min, int32_t max);
 } // namespace reference
 } // namespace validation
 } // namespace test