COMPMID-2794: Add support for QASYMM8_SIGNED in CLGEMMLowpOutputStage

Change-Id: I93ad3e5b9531ce1699214ff6e657a76ffdaacedd
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2396
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

10 files changed, 446 insertions, 18 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 78437beffb..d070d6a8c8 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -80,6 +80,7 @@
 #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/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.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/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..22ac8fae4a
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
@@ -0,0 +1,96 @@
+/*
+ * 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_CLGEMMLOWPQUANTIZEDOWNINT32TOINT8SCALEBYFIXEDPOINTKERNEL_H
+#define ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOINT8SCALEBYFIXEDPOINTKERNEL_H
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** OpenCL 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 CLGEMMLowpMatrixMultiplyKernel), 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 CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel : public ICLKernel
+{
+public:
+    /** Constructor */
+    CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel(const CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &operator=(const CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel(CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &operator=(CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &&) = 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: 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. Defaults to 0
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED. Defaults to 0
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     */
+    void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, 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 of @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
+     *
+     * @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[in] output Output tensor. Data type supported: Data type supported: 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
+     *
+     * @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_CLGEMMLOWPQUANTIZEDOWNINT32TOINT8SCALEBYFIXEDPOINTKERNEL_H */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
index 0e70223998..25fa142b21 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
@@ -21,15 +21,15 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H__
-#define __ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H__
+#ifndef ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H
+#define ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H
 
 #include "arm_compute/runtime/CL/ICLSimpleFunction.h"
 
 /** This file contains all available output stages for GEMMLowp on OpenCL.
  *
  *  In gemmlowp, the "output stage" is the process that takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyCore),
- *  and processes it to obtain the final ASYMM8 value.
+ *  and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value.
  *
  *  More information about the GEMMLowp output stage can be found at https://github.com/google/gemmlowp/blob/master/doc/output.md
  */
@@ -149,6 +149,67 @@ public:
     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
 };
 
+/** Basic function to execute CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint on OpenCL.
+ *
+ *  CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint depends on 3 parameters:
+ *
+ *  result_fixedpoint_multiplier, result_shift, result_offset_after_shift
+ *
+ * The final result is:
+ *
+ * (FixedPointMul(input[i][k], result_fixedpoint_multiplier) >> result_shift) + result_offset_after_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 OpenCL kernels:
+ *
+ * -# @ref CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
+ *
+ * @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 CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint : 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: 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                 Number of bits to shift right 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. Defaults to 0
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED. Defaults to 0
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     */
+    void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, 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 of @ref CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
+     *
+     * @param[in] input  Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
+     * @param[in] bias   Biases tensor. 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. Data type supported: Data type supported: QASYMM8_SIGNED
+     * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED. Defaults to 0
+     * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED. Defaults to 0
+     *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
+};
+
 /** Basic function to execute CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat on OpenCL.
  *
  *  This function calls the following OpenCL kernels:
diff --git a/docs/06_functions_list.dox b/docs/06_functions_list.dox
index 30b522bb2a..b6b94c4ade 100644
--- a/docs/06_functions_list.dox
+++ b/docs/06_functions_list.dox
@@ -93,6 +93,7 @@ namespace arm_compute
         - @ref NEGEMMInterleave4x4
         - @ref NEGEMMLowpQuantizeDownInt32ToUint8Scale
         - @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
+        - @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
         - @ref NEGEMMTranspose1xW
         - @ref NEHOGDetector
         - @ref NEMagnitude
@@ -298,6 +299,7 @@ namespace arm_compute
         - @ref CLGaussian3x3
         - @ref CLGEMMLowpQuantizeDownInt32ToUint8Scale
         - @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
+        - @ref CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
         - @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat
         - @ref CLMagnitude
         - @ref CLMeanStdDevNormalizationLayer
diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp
index 9754bebd18..47472a3dae 100644
--- a/src/core/CL/CLHelpers.cpp
+++ b/src/core/CL/CLHelpers.cpp
@@ -44,6 +44,7 @@ std::string get_cl_type_from_data_type(const DataType &dt)
         case DataType::S8:
         case DataType::QASYMM8_SIGNED:
         case DataType::QSYMM8:
+        case DataType::QASYMM8_SIGNED:
         case DataType::QSYMM8_PER_CHANNEL:
             return "char";
         case DataType::U16:
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index 47791fbe74..2a1c1561da 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -1824,11 +1824,14 @@ __kernel void gemmlowp_output_stage_quantize_down(TENSOR3D_DECLARATION(src),
  *  -# 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.
+ *  -# Clamp the resulting int32 values:
+ *      - to the [0..255] range and cast to QASYMM8.
+ *      - to the [-128..127] range and cast to QASYMM8_SIGNED.
  *
  * @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET_AFTER_SHIFT, -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 The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE
  * @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
  *
@@ -1888,17 +1891,18 @@ __kernel void gemmlowp_output_stage_quantize_down_fixedpoint(TENSOR3D_DECLARATIO
     // Add the offset terms to GEMM's result
     input_values += (int4)RESULT_OFFSET_AFTER_SHIFT;
 
-    uchar4 res = convert_uchar4_sat(input_values);
+    VEC_DATA_TYPE(OUTPUT_DATA_TYPE, 4)
+    res = CONVERT_SAT(input_values, VEC_DATA_TYPE(OUTPUT_DATA_TYPE, 4));
 
 #if defined(MIN_BOUND)
-    res = max(res, (uchar4)MIN_BOUND);
+    res = max(res, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, 4))MIN_BOUND);
 #endif // defined(MIN_BOUND)
 #if defined(MAX_BOUND)
-    res = min(res, (uchar4)MAX_BOUND);
+    res = min(res, (VEC_DATA_TYPE(OUTPUT_DATA_TYPE, 4))MAX_BOUND);
 #endif // defined(MAX_BOUND)
 
     // Store the result
-    vstore4(res, 0, dst_addr);
+    vstore4(res, 0, (__global OUTPUT_DATA_TYPE *)dst_addr);
 }
 #endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
 
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
new file mode 100644
index 0000000000..3de3182a57
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
@@ -0,0 +1,181 @@
+/*
+ * 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/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.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"
+
+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 > 127);
+    ARM_COMPUTE_RETURN_ERROR_ON(min < -128 || 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::QASYMM8_SIGNED);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    }
+
+    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::QASYMM8_SIGNED));
+
+    // 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
+
+CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel()
+    : _input(nullptr), _bias(nullptr), _output(nullptr)
+{
+}
+
+Status CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::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 CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
+                                                                         int result_fixedpoint_multiplier, int result_shift, int result_offset_after_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));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_and_configure_window(input->info(),
+                                                             (bias != nullptr) ? bias->info() : nullptr,
+                                                             output->info())
+                               .first);
+
+    _input  = input;
+    _bias   = bias;
+    _output = output;
+
+    // Set the arguments to pass at compile time
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DRESULT_OFFSET_AFTER_SHIFT=" + support::cpp11::to_string(result_offset_after_shift));
+    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("-DOUTPUT_DATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type()));
+    build_opts.add_option_if((min != -128) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+    build_opts.add_option_if((max != 127) && (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", 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 CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::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, lws_hint());
+    }
+    while(collapsed.slide_window_slice_3D(slice));
+}
+} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
index a98eae673b..a5b00d1e74 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -24,6 +24,7 @@
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
 
 #include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
@@ -75,15 +76,13 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 
     AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
 
-    bool window_changed = update_window_and_padding(win,
-                                                    input_access);
+    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()));
     }
 
@@ -122,8 +121,11 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const
 {
     // 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));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_and_configure_window(input->info(),
+                                                             (bias != nullptr) ? bias->info() : nullptr,
+                                                             output->info())
+                               .first);
 
     _input  = input;
     _bias   = bias;
@@ -134,6 +136,7 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const
     build_opts.add_option("-DRESULT_OFFSET_AFTER_SHIFT=" + support::cpp11::to_string(result_offset_after_shift));
     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("-DOUTPUT_DATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type()));
     build_opts.add_option_if((min != 0) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
     build_opts.add_option_if((max != 255) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
     build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
diff --git a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
index 020fbbe52c..9551fc7efb 100644
--- a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
@@ -25,6 +25,7 @@
 
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.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"
@@ -59,6 +60,21 @@ Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(const ITens
     return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, min, max);
 }
 
+void CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
+                                                                   int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift,
+                                                                   int min, int max)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel>();
+    k->configure(input, bias, output, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
+    _kernel = std::move(k);
+}
+
+Status CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
+                                                                    int min, int max)
+{
+    return CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(input, bias, output, min, max);
+}
+
 void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
                                                                float multiplier, int offset,
                                                                int min, int max)
diff --git a/tests/validation/CL/GEMMLowp.cpp b/tests/validation/CL/GEMMLowp.cpp
index 39543b174c..2890eb161b 100644
--- a/tests/validation/CL/GEMMLowp.cpp
+++ b/tests/validation/CL/GEMMLowp.cpp
@@ -211,9 +211,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpQuantizeDownInt32ToUint8ScaleFixture,
 }
 TEST_SUITE_END() // BoundedReLu
 TEST_SUITE_END() // QuantizeDownInt32ToUint8Scale
-
 TEST_SUITE(QuantizeDownInt32ToUint8ScaleByFixedPoint)
-
 const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
                                                                     2)
                                                                     * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0) * framework::dataset::make("max", 0) * framework::dataset::make("addBias", { false, true });
@@ -221,12 +219,10 @@ const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_cases = framework::d
 const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
                                                                          2)
                                                                          * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 174) * framework::dataset::make("addBias", { false, true });
-
 using CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture =
     GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture<CLTensor, CLAccessor, CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint>;
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
-                                                                   quantize_down_int32_to_uint8_scale_by_fixedpoint_cases),
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_by_fixedpoint_cases),
                shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias)
 {
     TensorShape shape_bias(shape[0]);
@@ -297,6 +293,73 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedP
 }
 TEST_SUITE_END() // BoundedReLu
 TEST_SUITE_END() // QuantizeDownInt32ToUint8ScaleByFixedPoint
+TEST_SUITE(QuantizeDownInt32ToInt8ScaleByFixedPoint)
+const auto quantize_down_int32_to_int8_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, 2)
+                                                                   * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0) * framework::dataset::make("max", 0) * framework::dataset::make("addBias", { false, true });
+
+const auto quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, 2)
+                                                                        * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", -128, -126) * framework::dataset::make("max", 110, 112) * framework::dataset::make("addBias", { false, true });
+using CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture =
+    GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture<CLTensor, CLAccessor, CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint>;
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_by_fixedpoint_cases),
+               shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias)
+{
+    TensorShape shape_bias(shape[0]);
+
+    // Create tensors
+    CLTensor in   = create_tensor<CLTensor>(shape, DataType::S32);
+    CLTensor bias = create_tensor<CLTensor>(shape_bias, DataType::S32);
+    CLTensor out  = create_tensor<CLTensor>(shape, DataType::QASYMM8_SIGNED);
+
+    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
+    CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint output_stage;
+    output_stage.configure(&in, add_bias ? &bias : nullptr, &out, result_fixedpoint_multiplier, result_shift, result_offset_after_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 = PaddingCalculator(shape.x(), 4).required_padding();
+    validate(in.info()->padding(), padding);
+    validate(out.info()->padding(), padding);
+
+    if(add_bias)
+    {
+        validate(bias.info()->padding(), padding);
+    }
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int8_scale_by_fixedpoint_cases))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+TEST_SUITE(BoundedReLu)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+TEST_SUITE_END() // BoundedReLu
+TEST_SUITE_END() // QuantizeDownInt32ToInt8ScaleByFixedPoint
 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,