COMPMID-2410: Create a new GEMMLowpQuantizeDownInt32ToInt16ScaleKernel for CL

Change-Id: Iab74b72f7adf712a1baf16aab916ea7c8d2bf92f
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1497
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

8 files changed, 499 insertions, 2 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 8fbc4770b0..cd9eadc1cd 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -81,6 +81,7 @@
 #include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..2bd2bb6afb
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
@@ -0,0 +1,94 @@
+/*
+ * 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_CLGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H__
+#define __ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** CL 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 CLGEMMLowpMatrixMultiplyKernel), 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 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel : public ICLKernel
+{
+public:
+    /** Constructor */
+    CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(const CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(const CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = 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 ICLTensor *input, const ICLTensor *bias, ICLTensor *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 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+     *
+     * @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, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor *_input;
+    const ICLTensor *_bias;
+    ICLTensor       *_output;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
index cfd1f08519..0e70223998 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -188,5 +188,63 @@ public:
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
 };
+/** Basic function to execute CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint on OpenCL.
+ *
+ *  CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint 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 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+ *
+ * @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 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint : public ICLSimpleFunction
+{
+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 ICLTensor *input, const ICLTensor *bias, ICLTensor *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 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint
+     *
+     * @param[in] input  Input tensor info. It is the output of @ref CLGEMMLowpMatrixMultiplyCore 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_CLGEMMLOWPOUTPUTSTAGE_H__ */
\ No newline at end of file
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 36d8bed5b9..8b64b1f20e 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -342,6 +342,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemmlowp_offset_contribution_quantize_down_fixedpoint", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down_fixedpoint", "gemmlowp.cl" },
+    { "gemmlowp_output_stage_quantize_down_fixedpoint_qsymm16", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down_float", "gemmlowp.cl" },
     { "generate_proposals_compute_all_anchors", "generate_proposals.cl" },
     { "harris_score_3x3", "harris_corners.cl" },
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index 65c31efe2b..4b869554c5 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -2861,6 +2861,89 @@ __kernel void gemmlowp_output_stage_quantize_down_fixedpoint(TENSOR3D_DECLARATIO
 }
 #endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
 
+#if defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
+
+/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM16
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), 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
+ *  -# Add offset to each result
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [-32768..32767] range and cast to QSYMM16.
+ *
+ * @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_FIXEDPOINT_MULTIPLIER and -DRESULT_SHIFT
+ *
+ * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time
+ * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.
+ *       These values can be used to implement "rectified linear unit" activation functions
+ *
+ * @param[in]  src_ptr                              Pointer to the source tensor. Supported data type: S32
+ * @param[in]  src_stride_x                         Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                           src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                         Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                           src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                         Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                           src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes    The offset of the first element in the source tensor
+ * @param[in]  biases_ptr                           (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr
+ * @param[in]  biases_stride_x                      (Optional) Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                        (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor
+ * @param[out] dst_ptr                              Pointer to the destination tensor Supported data type: QASYMM8
+ * @param[in]  dst_stride_x                         Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  dst_step_x                           dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                         Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                           dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                         Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                           src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes    The offset of the first element in the destination tensor
+ */
+__kernel void gemmlowp_output_stage_quantize_down_fixedpoint_qsymm16(TENSOR3D_DECLARATION(src),
+#if defined(ADD_BIAS)
+                                                                     VECTOR_DECLARATION(biases),
+#endif // defined(ADD_BIAS)
+                                                                     TENSOR3D_DECLARATION(dst))
+{
+    // Compute source and destination addresses
+    int x = get_global_id(0) * 4;
+    int y = get_global_id(1);
+    int z = get_global_id(2);
+
+    __global short *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(int) + y * src_stride_y + z * src_stride_z;
+
+    __global short *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * 2 + y * dst_stride_y + z * dst_stride_z;
+
+    int4 input_values = vload4(0, (__global int *)src_addr);
+
+#if defined(ADD_BIAS)
+    // Add bias
+    __global short *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int);
+
+    int4 biases_values = vload4(0, (__global int *)bias_addr);
+    input_values += (int4)biases_values;
+#endif // defined(ADD_BIAS)
+
+    // Multiply by result_mult_int and shift
+    input_values = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(input_values, RESULT_FIXEDPOINT_MULTIPLIER, RESULT_SHIFT, 4);
+
+    short4 res = convert_short4_sat(input_values);
+
+#if defined(MIN_BOUND)
+    res = max(res, (short4)MIN_BOUND);
+#endif // defined(MIN_BOUND)
+#if defined(MAX_BOUND)
+    res = min(res, (short4)MAX_BOUND);
+#endif // defined(MAX_BOUND)
+
+    // Store the result
+    vstore4(res, 0, dst_addr);
+}
+#endif // defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
+
 #if defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)
 /** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
  *
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
new file mode 100644
index 0000000000..557e82dc50
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
@@ -0,0 +1,180 @@
+/*
+ * Copyright (c) 2017-2019 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/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+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 *bias, ITensorInfo *output)
+{
+    constexpr unsigned int num_elems_processed_per_iteration = 4;
+
+    // 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(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+
+    bool window_changed = update_window_and_padding(win, input_access);
+
+    if(output->total_size() != 0)
+    {
+        Window                 win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+        AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
+        window_changed = window_changed || update_window_and_padding(win_out, output_result_access);
+
+        output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    }
+
+    if(bias != nullptr)
+    {
+        AccessWindowStatic bias_access(bias, 0, 0, ceil_to_multiple(bias->dimension(0), num_elems_processed_per_iteration), bias->tensor_shape()[1]);
+        window_changed = window_changed || update_window_and_padding(win, bias_access);
+    }
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+class Coordinates;
+} // namespace arm_compute
+
+CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel()
+    : _input(nullptr), _bias(nullptr), _output(nullptr)
+{
+}
+
+Status CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::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(),
+                                                              (bias != nullptr) ? bias->clone().get() : nullptr,
+                                                              output->clone().get())
+                                .first);
+
+    return Status{};
+}
+
+void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *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;
+
+    // Set the arguments to pass at compile time
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DRESULT_FIXEDPOINT_MULTIPLIER=" + support::cpp11::to_string(result_fixedpoint_multiplier));
+    build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(result_shift));
+    build_opts.add_option_if((min != -32768) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+    build_opts.add_option_if((max != 32767) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
+    build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_output_stage_quantize_down_fixedpoint_qsymm16", build_opts.options()));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+}
+
+void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    // Create input window
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+    Window slice     = collapsed.first_slice_window_3D();
+
+    // Setup bias slice
+    unsigned int idx1 = num_arguments_per_3D_tensor();
+    if(_bias != nullptr)
+    {
+        Window biases_slice(slice);
+        biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
+        biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
+        add_1D_tensor_argument(idx1, _bias, biases_slice);
+    }
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx1, _output, slice);
+        enqueue(queue, *this, slice);
+    }
+    while(collapsed.slide_window_slice_3D(slice));
+}
diff --git a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
index f1282cbde9..020fbbe52c 100644
--- a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,7 @@
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
 
 #include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
@@ -72,4 +73,20 @@ Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat::validate(const ITensorInf
 {
     return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::validate(input, bias, output, min, max);
 }
+
+void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
+                                                                    int result_fixedpoint_multiplier, int result_shift,
+                                                                    int min, int max)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel>();
+    k->configure(input, bias, output, result_fixedpoint_multiplier, result_shift, min, max);
+    _kernel = std::move(k);
+}
+
+Status CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
+                                                                     int min, int max)
+{
+    return CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(input, bias, output, min, max);
+}
+
 } // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/CL/GEMMLowp.cpp b/tests/validation/CL/GEMMLowp.cpp
index efefbd645b..b8dfc030a2 100644
--- a/tests/validation/CL/GEMMLowp.cpp
+++ b/tests/validation/CL/GEMMLowp.cpp
@@ -295,7 +295,70 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedP
 }
 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 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture =
+    GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<CLTensor, CLAccessor, CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>;
+
+// *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),
+                                             TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Invalid min and max
+                                             TensorInfo(TensorShape(21U, 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(21U), 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 =  CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(&a_info.clone()->set_is_resizable(true),
+                                                                                 &b_info.clone()->set_is_resizable(true),
+                                                                                 &output_info.clone()->set_is_resizable(true),
+                                                                                 min,
+                                                                                 max);
+    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int16_scale_by_fixedpoint_cases))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE(BoundedReLu)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // BoundedReLu
+TEST_SUITE_END() // QuantizeDownInt32ToInt16ScaleByFixedPoint
 TEST_SUITE_END() // OutputStage
 TEST_SUITE_END() // GEMMLowp
 TEST_SUITE_END() // CL