COMPMID-2966 Add support for QASYMM8_SIGNED in NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel

Signed-off-by: Luca Foschiani <luca.foschiani@arm.com>
Change-Id: Ia8692f8fda16fa3b73f343e4b5b1b55e14403225
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2750
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

13 files changed, 557 insertions, 393 deletions

diff --git a/Android.bp b/Android.bp
index 0d5c9e949d..0cb0b7770e 100644
--- a/Android.bp
+++ b/Android.bp
@@ -281,10 +281,10 @@ cc_library_static {
         "src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp",
         "src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp",
         "src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp",
+        "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.cpp",
         "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp",
         "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp",
         "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp",
-        "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp",
         "src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp",
         "src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp",
         "src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.cpp",
diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleKernel.h
index f9599b5a0e..3378359d29 100644
--- a/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleKernel.h
@@ -64,7 +64,7 @@ public:
      * @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/QASYMM8_SIGNED
-     * @param[in]  output_stage Output stage info. Used to pass the quantized output data type
+     * @param[in]  output_stage GEMMLowp output stage metadata.
      */
     void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const GEMMLowpOutputStageInfo *output_stage);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpQuantizeDownInt32ScaleKernel
@@ -73,7 +73,7 @@ public:
      * @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/QASYMM8_SIGNED
-     * @param[in] output_stage Output stage info. Used to pass the quantized output data type
+     * @param[in] output_stage GEMMLowp output stage metadata.
      *
      * @return a status
      */
diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 5daad34468..d9f8f00c0b 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2019 ARM Limited.
+ * Copyright (c) 2016-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -79,10 +79,10 @@
 #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/NEGEMMLowpQuantizeDownInt32ScaleKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.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"
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAdditionKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.h
new file mode 100644
index 0000000000..b4a1419c9b
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.h
@@ -0,0 +1,112 @@
+/*
+ * Copyright (c) 2020 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_NEGEMMLOWPQUANTIZEDOWNINT32SCALEKERNEL_H
+#define ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32SCALEKERNEL_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 QASYMM8/QASYMM8_SIGNED
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Add offset terms to final result
+ *  -# Multiply each entry of result by result_mult_int
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Shift the int32 accumulator by result_shift
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values:
+ *  -#  -to the [0..255] range and cast to QASYMM8.
+ *  -#  -to the [-128..127] range and cast to QASYMM8_SIGNED.
+ *
+ */
+class NEGEMMLowpQuantizeDownInt32ScaleKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEGEMMLowpQuantizeDownInt32ScaleKernel";
+    }
+    /** Constructor */
+    NEGEMMLowpQuantizeDownInt32ScaleKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ScaleKernel(const NEGEMMLowpQuantizeDownInt32ScaleKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ScaleKernel &operator=(const NEGEMMLowpQuantizeDownInt32ScaleKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ScaleKernel(NEGEMMLowpQuantizeDownInt32ScaleKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ScaleKernel &operator=(NEGEMMLowpQuantizeDownInt32ScaleKernel &&) = 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/QASYMM8_SIGNED
+     * @param[out] output_stage GEMMLowp output stage metadata.
+     */
+    void configure(const ITensor *input, const ITensor *bias, ITensor *output, const GEMMLowpOutputStageInfo *output_stage);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ScaleKernel
+     *
+     * @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/QASYMM8_SIGNED
+     * @param[out] output_stage GEMMLowp output stage metadata.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    /** Template function to run the NEGEMMLowpQuantizeDownInt32ScaleKernel
+     *
+     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <typename T>
+    void run(const Window &window);
+
+    /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ScaleKernel functions
+     *
+     * @param[in] window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ScaleKernel::*)(const Window &window);
+
+    QuantizeDownFunctionPtr        _func;
+    const ITensor                 *_input;
+    const ITensor                 *_bias;
+    ITensor                       *_output;
+    const GEMMLowpOutputStageInfo *_output_stage;
+    bool                           _is_bounded_relu;
+};
+} // namespace arm_compute
+
+#endif /* ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32SCALEKERNEL_H */
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h
deleted file mode 100644
index 14cc383014..0000000000
--- a/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h
+++ /dev/null
@@ -1,120 +0,0 @@
-/*
- * 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.
- */
-#ifndef ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEKERNEL_H
-#define ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEKERNEL_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 QASYMM8
- *
- * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
- * The following computations will be performed by the kernel:
- *
- *  -# Add offset terms to final result
- *  -# Multiply each entry of result by result_mult_int
- *  -# Add bias to final result if bias tensor is not a nullptr
- *  -# Shift the int32 accumulator by result_shift
- *  -# Clamp the value between the specified min and max bounds
- *  -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
- *
- */
-class NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel : public INEKernel
-{
-public:
-    const char *name() const override
-    {
-        return "NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel";
-    }
-    /** Constructor */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel(NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &operator=(NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &&) = 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
-     * @param[in]  result_offset   Offset to be added to each element of the input matrix
-     * @param[in]  result_mult_int 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 before converting back to QASYMM8
-     * @param[in]  min             (Optional) Min value used to saturate down the output result before converting back to QASYMM8
-     * @param[in]  max             (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
-     *                             Along with @p min, this value can be used to implement "rectified linear unit" activation functions
-     */
-    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_offset, int result_mult_int, int result_shift, int min = 0, int max = 0);
-    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel
-     *
-     * @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
-     * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8
-     * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
-     *                   Along with @p min, this value can be used to implement "rectified linear unit" activation functions
-     *
-     * @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 NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel
-     *
-     * @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 NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel functions
-     *
-     * @param[in] window Region on which to execute the kernel.
-     */
-    using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::*)(const Window &window);
-
-    QuantizeDownFunctionPtr _func;
-    const ITensor          *_input;
-    const ITensor          *_bias;
-    ITensor                *_output;
-    int                     _result_offset;
-    int                     _result_mult_int;
-    int                     _result_shift;
-    int                     _min;
-    int                     _max;
-};
-} // namespace arm_compute
-
-#endif /* ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEKERNEL_H */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
index 283b052917..cbdc788c0a 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
@@ -51,7 +51,7 @@ class ITensor;
  *
  *  This function calls the following NEON kernels:
  *
- * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel
+ * -# @ref NEGEMMLowpQuantizeDownInt32ScaleKernel
  *
  * @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
@@ -72,6 +72,7 @@ public:
      * @param[in]  max             (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
      *                             Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to the maximum possible 32-bit signed integer.
      */
+    ARM_COMPUTE_DEPRECATED_REL(20.05)
     void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_offset, int result_mult_int, int result_shift, int min = std::numeric_limits<int32_t>::lowest(),
                    int max = std::numeric_limits<int32_t>::max());
     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToUint8Scale
@@ -86,6 +87,7 @@ public:
      *
      * @return a status
      */
+    ARM_COMPUTE_DEPRECATED_REL(20.05)
     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = std::numeric_limits<int32_t>::lowest(), int max = std::numeric_limits<int32_t>::max());
 };
 
@@ -273,7 +275,7 @@ public:
  *
  *  This function calls the following NEON kernels:
  *
- * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel
+ * -# @ref NEGEMMLowpQuantizeDownInt32ScaleKernel
  * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
  * -# @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
  * -# @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
diff --git a/docs/00_introduction.dox b/docs/00_introduction.dox
index d3ec24d743..67b879c37b 100644
--- a/docs/00_introduction.dox
+++ b/docs/00_introduction.dox
@@ -855,7 +855,6 @@ v17.12 Public major release
     - @ref NEDepthwiseConvolutionLayer3x3Kernel / NEDepthwiseIm2ColKernel / @ref NEGEMMMatrixVectorMultiplyKernel / NEDepthwiseVectorToTensorKernel / @ref NEDepthwiseConvolutionLayer
     - @ref NEGEMMLowpOffsetContributionKernel / @ref NEGEMMLowpMatrixAReductionKernel / @ref NEGEMMLowpMatrixBReductionKernel / @ref NEGEMMLowpMatrixMultiplyCore
     - @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel / @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
-    - @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel / @ref NEGEMMLowpQuantizeDownInt32ToUint8Scale
     - NEWinogradLayer / NEWinogradLayerKernel
 
  - New OpenCL kernels / functions
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.cpp
index a68e4e7efb..80ba2aff93 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 ARM Limited.
+ * Copyright (c) 2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,29 +21,32 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.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/wrapper/wrapper.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 
 #include <arm_neon.h>
 #include <cstddef>
 #include <cstdint>
 
-using namespace arm_compute;
-
-namespace
+namespace arm_compute
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_max_bound > std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type)));
+    ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_min_bound < std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))
+                                || output_stage->gemmlowp_min_bound > output_stage->gemmlowp_max_bound);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -55,46 +58,17 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
 
     if(output->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+        if(output->data_type() != output_stage->output_data_type && (output_stage->output_data_type == DataType::QASYMM8 || output_stage->output_data_type == DataType::QASYMM8_SIGNED))
+        {
+            ARM_COMPUTE_RETURN_ERROR_MSG("Mismatching data types");
+        }
+
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
     }
 
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output)
-{
-    // Note: This kernel performs 16 elements per iteration.
-    // However, since we use a left-over for loop, we cannot have any read or write out of memory
-    // For this reason num_elems_processed_per_iteration is set to 1
-    constexpr unsigned int num_elems_processed_per_iteration = 1;
-
-    // Configure kernel window
-    Window win = calculate_max_window(*output, 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)
-    {
-        AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
-        window_changed = window_changed || update_window_and_padding(win, output_result_access);
-
-        output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
-    }
-
-    if(bias != nullptr)
-    {
-        AccessWindowStatic bias_access(bias, 0, 0, bias->dimension(0), bias->dimension(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);
-}
-
 inline void scale_input(int32x4x4_t &in_s32, int32x4_t result_offset_s32, int32_t result_mult_int)
 {
     // Add the offset terms to GEMM's result
@@ -110,23 +84,32 @@ inline void scale_input(int32x4x4_t &in_s32, int32x4_t result_offset_s32, int32_
     in_s32.val[3] = vmulq_n_s32(in_s32.val[3], result_mult_int);
 }
 
-template <bool    is_bounded_relu>
-inline uint8x16_t finalize_quantization(int32x4x4_t &in_s32, int32x4_t result_shift_s32, uint8x16_t min_u8, uint8x16_t max_u8)
+template <typename T>
+inline typename std::enable_if<std::is_same<T, uint8_t>::value,
+       typename wrapper::traits::neon_vector<T, 16>::type>::type
+       convert_to_8bit(const int16x8x2_t in_s16)
+{
+    return wrapper::vcombine(wrapper::vqmovun(in_s16.val[0]), wrapper::vqmovun(in_s16.val[1]));
+}
+
+template <typename T>
+inline typename std::enable_if<std::is_same<T, int8_t>::value,
+       typename wrapper::traits::neon_vector<T, 16>::type>::type
+       convert_to_8bit(const int16x8x2_t in_s16)
 {
-    const static int32x4_t zero_s32 = vdupq_n_s32(0);
+    return wrapper::vcombine(wrapper::vqmovn(in_s16.val[0]), wrapper::vqmovn(in_s16.val[1]));
+}
 
+template <typename T>
+inline typename wrapper::traits::neon_vector<T, 16>::type finalize_quantization(int32x4x4_t &in_s32, int32x4_t result_shift_s32, typename wrapper::traits::neon_vector<T, 16>::type min,
+                                                                                typename wrapper::traits::neon_vector<T, 16>::type max)
+{
     // Shift final result (negative value shift right)
     in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32);
     in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32);
     in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32);
     in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32);
 
-    // Saturate negative values
-    in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32);
-    in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32);
-    in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32);
-    in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32);
-
     // Convert S32 to S16
     const int16x8x2_t in_s16 =
     {
@@ -136,38 +119,33 @@ inline uint8x16_t finalize_quantization(int32x4x4_t &in_s32, int32x4_t result_sh
         }
     };
 
-    // Convert S16 to U8
-    uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1]));
+    // Convert S16 to S8 or U8
+    typename wrapper::traits::neon_vector<T, 16>::type out = convert_to_8bit<T>(in_s16);
 
-    if(is_bounded_relu)
-    {
-        out_u8 = vmaxq_u8(out_u8, min_u8);
-        out_u8 = vminq_u8(out_u8, max_u8);
-    }
+    out = wrapper::vmax(out, min);
+    out = wrapper::vmin(out, max);
 
-    return out_u8;
+    return out;
 }
-} // namespace
 
-namespace arm_compute
-{
 class Coordinates;
-} // namespace arm_compute
 
-template <bool is_bounded_relu>
-void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window)
+template <typename T>
+void NEGEMMLowpQuantizeDownInt32ScaleKernel::run(const Window &window)
 {
-    const int32x4_t  result_offset_s32 = vdupq_n_s32(_result_offset);
-    const int32x4_t  result_shift_s32  = vdupq_n_s32(-_result_shift);
-    const uint8x16_t min_u8            = vdupq_n_u8(static_cast<uint8_t>(_min));
-    const uint8x16_t max_u8            = vdupq_n_u8(static_cast<uint8_t>(_max));
+    using VectorType = typename wrapper::traits::neon_vector<T, 16>::type;
 
-    ARM_COMPUTE_UNUSED(min_u8);
-    ARM_COMPUTE_UNUSED(max_u8);
+    const int32x4_t result_offset_s32 = vdupq_n_s32(_output_stage->gemmlowp_offset);
+    const int32x4_t result_shift_s32  = vdupq_n_s32(-_output_stage->gemmlowp_shift);
+    const int       window_step_x     = 16;
+    const auto      window_start_x    = static_cast<int>(window.x().start());
+    const auto      window_end_x      = static_cast<int>(window.x().end());
 
-    const int  window_step_x  = 16;
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
+    const int clamp_min = (_is_bounded_relu) ? _output_stage->gemmlowp_min_bound : std::numeric_limits<T>::lowest();
+    const int clamp_max = (_is_bounded_relu) ? _output_stage->gemmlowp_max_bound : std::numeric_limits<T>::max();
+
+    VectorType min = wrapper::vdup_n(static_cast<T>(clamp_min), wrapper::traits::vector_128_tag{});
+    VectorType max = wrapper::vdup_n(static_cast<T>(clamp_max), wrapper::traits::vector_128_tag{});
 
     Window win(window);
     win.set(Window::DimX, Window::Dimension(0, 1, 1));
@@ -215,9 +193,9 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window)
                 in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
 
                 // Add the offset terms to GEMM's result and multiply by result_mult_int
-                scale_input(in_s32, result_offset_s32, _result_mult_int);
+                scale_input(in_s32, result_offset_s32, _output_stage->gemmlowp_multiplier);
 
-                vst1q_u8(out.ptr() + x, finalize_quantization<is_bounded_relu>(in_s32, result_shift_s32, min_u8, max_u8));
+                wrapper::vstore(reinterpret_cast<T *>(out.ptr() + x), finalize_quantization<T>(in_s32, result_shift_s32, min, max));
             }
 
             // Compute left-over elements
@@ -227,17 +205,10 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window)
                 int       in_value   = *(reinterpret_cast<const int *>(in.ptr()) + x);
 
                 // Quantize
-                in_value = ((in_value + bias_value + _result_offset) * _result_mult_int) >> _result_shift;
+                in_value = ((in_value + bias_value + _output_stage->gemmlowp_offset) * _output_stage->gemmlowp_multiplier) >> _output_stage->gemmlowp_shift;
 
-                // Finalize and store the result
-                if(is_bounded_relu)
-                {
-                    *(out.ptr() + x) = static_cast<uint8_t>(std::max(_min, std::min(_max, in_value)));
-                }
-                else
-                {
-                    *(out.ptr() + x) = static_cast<uint8_t>(std::max(0, std::min(255, in_value)));
-                }
+                // Store the result
+                *(out.ptr() + x) = static_cast<T>(utility::clamp<int>(in_value, clamp_min, clamp_max));
             }
         },
         in, bias, out);
@@ -261,9 +232,9 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window)
                 };
 
                 // Add the offset terms to GEMM's result and multiply by result_mult_int
-                scale_input(in_s32, result_offset_s32, _result_mult_int);
+                scale_input(in_s32, result_offset_s32, _output_stage->gemmlowp_multiplier);
 
-                vst1q_u8(out.ptr() + x, finalize_quantization<is_bounded_relu>(in_s32, result_shift_s32, min_u8, max_u8));
+                wrapper::vstore(reinterpret_cast<T *>(out.ptr() + x), finalize_quantization<T>(in_s32, result_shift_s32, min, max));
             }
 
             // Compute left-over elements
@@ -272,74 +243,74 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window)
                 int in_value = *(reinterpret_cast<const int *>(in.ptr()) + x);
 
                 // Quantize
-                in_value = ((in_value + _result_offset) * _result_mult_int) >> _result_shift;
+                in_value = ((in_value + _output_stage->gemmlowp_offset) * _output_stage->gemmlowp_multiplier) >> _output_stage->gemmlowp_shift;
 
-                // Finalize and store the result
-                if(is_bounded_relu)
-                {
-                    *(out.ptr() + x) = static_cast<uint8_t>(std::max(_min, std::min(_max, in_value)));
-                }
-                else
-                {
-                    *(out.ptr() + x) = static_cast<uint8_t>(std::max(0, std::min(255, in_value)));
-                }
+                // Store the result
+                *(out.ptr() + x) = static_cast<T>(utility::clamp<int>(in_value, clamp_min, clamp_max));
             }
         },
         in, out);
     }
 }
 
-NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel()
-    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_offset(0), _result_mult_int(0), _result_shift(0), _min(0), _max(0)
+NEGEMMLowpQuantizeDownInt32ScaleKernel::NEGEMMLowpQuantizeDownInt32ScaleKernel()
+    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _output_stage(nullptr), _is_bounded_relu(false)
 {
 }
 
-void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_offset, int result_mult_int, int result_shift, int min, int max)
+void NEGEMMLowpQuantizeDownInt32ScaleKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, const GEMMLowpOutputStageInfo *output_stage)
 {
     // Perform validate step
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, output_stage);
 
     // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(DataType::QASYMM8));
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(output_stage->output_data_type));
 
     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_offset   = result_offset;
-    _result_mult_int = result_mult_int;
-    _result_shift    = result_shift;
-    _min             = min;
-    _max             = max;
+                                                  output_stage));
+
+    _input        = input;
+    _bias         = bias;
+    _output       = output;
+    _output_stage = output_stage;
 
     // 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);
-    INEKernel::configure(win_config.second);
+    Window      win = calculate_max_window(*input->info(), Steps());
+    Coordinates coord;
+    coord.set_num_dimensions(output->info()->num_dimensions());
+    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
 
     // Check if we need to clamp the result using min and max
-    const bool is_bounded_relu = !(min <= 0 && max >= 255);
-    _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run<false>;
+    _is_bounded_relu = ((_output_stage->gemmlowp_min_bound != _output_stage->gemmlowp_max_bound)
+                        && !(_output_stage->gemmlowp_min_bound == std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))
+                             && _output_stage->gemmlowp_max_bound == std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))));
+    if(_output_stage->output_data_type == DataType::QASYMM8)
+    {
+        _func = &NEGEMMLowpQuantizeDownInt32ScaleKernel::run<uint8_t>;
+    }
+    else if(_output_stage->output_data_type == DataType::QASYMM8_SIGNED)
+    {
+        _func = &NEGEMMLowpQuantizeDownInt32ScaleKernel::run<int8_t>;
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Data type not supported");
+    }
 }
 
-Status NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+Status NEGEMMLowpQuantizeDownInt32ScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
 {
     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);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, output_stage));
 
     return Status{};
 }
 
-void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window, const ThreadInfo &info)
+void NEGEMMLowpQuantizeDownInt32ScaleKernel::run(const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
@@ -347,3 +318,4 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window, co
 
     (this->*_func)(window);
 }
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
index fbd1820098..2114d39866 100644
--- a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
@@ -156,22 +156,9 @@ void CLGEMMLowpOutputStage::configure(const ICLTensor *input, const ICLTensor *b
         }
         case GEMMLowpOutputStageType::QUANTIZE_DOWN:
         {
-            switch(info.output_data_type)
-            {
-                case DataType::QASYMM8:
-                case DataType::QASYMM8_SIGNED:
-                {
-                    auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleKernel>();
-                    k->configure(input, bias, output, &info);
-                    _kernel = std::move(k);
-                    break;
-                }
-                default:
-                {
-                    ARM_COMPUTE_ERROR("Unsupported output data type.");
-                    break;
-                }
-            }
+            auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleKernel>();
+            k->configure(input, bias, output, &info);
+            _kernel = std::move(k);
             break;
         }
         case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT:
@@ -206,22 +193,9 @@ Status CLGEMMLowpOutputStage::validate(const ITensorInfo *input, const ITensorIn
             }
         }
         case GEMMLowpOutputStageType::QUANTIZE_DOWN:
-        {
-            switch(output->data_type())
-            {
-                case DataType::QASYMM8:
-                case DataType::QASYMM8_SIGNED:
-                {
-                    return CLGEMMLowpQuantizeDownInt32ScaleKernel::validate(input, bias, output, &info);
-                }
-                default:
-                    return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
-            }
-        }
+            return CLGEMMLowpQuantizeDownInt32ScaleKernel::validate(input, bias, output, &info);
         case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT:
-        {
             return CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::validate(input, bias, output, &info);
-        }
         default:
             return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported GEMMLowpOutputStage type.");
     }
diff --git a/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp b/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp
index 42d2ffce58..43ca7b3fbb 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp
@@ -24,10 +24,10 @@
 #include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
 
 #include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ScaleKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
 #include "arm_compute/core/Validate.h"
 #include "support/MemorySupport.h"
 
@@ -35,14 +35,25 @@ 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>();
-    k->configure(input, bias, output, result_offset, result_mult_int, result_shift, min, max);
+    GEMMLowpOutputStageInfo info = GEMMLowpOutputStageInfo();
+    info.gemmlowp_offset         = result_offset;
+    info.gemmlowp_multiplier     = result_mult_int;
+    info.gemmlowp_shift          = result_shift;
+    info.gemmlowp_min_bound      = min;
+    info.gemmlowp_max_bound      = max;
+
+    auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ScaleKernel>();
+    k->configure(input, bias, output, &info);
     _kernel = std::move(k);
 }
 
 Status NEGEMMLowpQuantizeDownInt32ToUint8Scale::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
-    return NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::validate(input, bias, output, min, max);
+    GEMMLowpOutputStageInfo info = GEMMLowpOutputStageInfo();
+    info.gemmlowp_min_bound      = min;
+    info.gemmlowp_max_bound      = max;
+
+    return NEGEMMLowpQuantizeDownInt32ScaleKernel::validate(input, bias, output, &info);
 }
 
 void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
@@ -89,53 +100,63 @@ void NEGEMMLowpOutputStage::configure(const ITensor *input, const ITensor *bias,
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpOutputStage::validate(input->info(), bias != nullptr ? bias->info() : nullptr, output->info(), info));
 
-    if(info.type == GEMMLowpOutputStageType::QUANTIZE_DOWN)
+    switch(info.type)
     {
-        switch(output->info()->data_type())
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
         {
-            case DataType::QASYMM8:
+            switch(info.output_data_type)
             {
-                auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel>();
-                k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                _kernel = std::move(k);
-                break;
+                case DataType::QASYMM8:
+                {
+                    auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel>();
+                    k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                    _kernel = std::move(k);
+                    break;
+                }
+                case DataType::QASYMM8_SIGNED:
+                {
+                    auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel>();
+                    k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                    _kernel = std::move(k);
+                    break;
+                }
+                case DataType::QSYMM16:
+                {
+                    auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel>();
+                    k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                    _kernel = std::move(k);
+                    break;
+                }
+                default:
+                {
+                    ARM_COMPUTE_ERROR("Unsupported output data type.");
+                    break;
+                }
             }
-            default:
-                ARM_COMPUTE_ERROR("Unsupported output data type.");
+            break;
         }
-    }
-    else if(info.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
-    {
-        switch(output->info()->data_type())
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
         {
-            case DataType::QASYMM8:
-            {
-                auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel>();
-                k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                _kernel = std::move(k);
-                break;
-            }
-            case DataType::QASYMM8_SIGNED:
-            {
-                auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel>();
-                k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                _kernel = std::move(k);
-                break;
-            }
-            case DataType::QSYMM16:
+            switch(info.output_data_type)
             {
-                auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel>();
-                k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-                _kernel = std::move(k);
-                break;
+                case DataType::QASYMM8:
+                case DataType::QASYMM8_SIGNED:
+                {
+                    auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpQuantizeDownInt32ScaleKernel>();
+                    k->configure(input, bias, output, &info);
+                    _kernel = std::move(k);
+                    break;
+                }
+                default:
+                {
+                    ARM_COMPUTE_ERROR("Unsupported output data type.");
+                    break;
+                }
             }
-            default:
-                ARM_COMPUTE_ERROR("Unsupported output data type.");
+            break;
         }
-    }
-    else
-    {
-        ARM_COMPUTE_ERROR("Unsupported output stage quantization type.");
+        default:
+            ARM_COMPUTE_ERROR("Unsupported GEMMLowpOutputStage type.");
     }
 }
 
@@ -147,29 +168,35 @@ Status NEGEMMLowpOutputStage::validate(const ITensorInfo *input, const ITensorIn
 
     ARM_COMPUTE_RETURN_ERROR_ON((info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN) && (info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT));
 
-    if(info.type == GEMMLowpOutputStageType::QUANTIZE_DOWN)
+    switch(info.type)
     {
-        switch(output->data_type())
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
         {
-            case DataType::QASYMM8:
-                return NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-            default:
-                return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
+            switch(output->data_type())
+            {
+                case DataType::QASYMM8:
+                    return NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                case DataType::QASYMM8_SIGNED:
+                    return NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                case DataType::QSYMM16:
+                    return NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+                default:
+                    return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
+            }
         }
-    }
-    else
-    {
-        switch(output->data_type())
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
         {
-            case DataType::QASYMM8:
-                return NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-            case DataType::QASYMM8_SIGNED:
-                return NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-            case DataType::QSYMM16:
-                return NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
-            default:
-                return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
+            switch(output->data_type())
+            {
+                case DataType::QASYMM8:
+                case DataType::QASYMM8_SIGNED:
+                    return NEGEMMLowpQuantizeDownInt32ScaleKernel::validate(input, bias, output, &info);
+                default:
+                    return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
+            }
         }
+        default:
+            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported GEMMLowpOutputStage type.");
     }
 }
 } // namespace arm_compute
diff --git a/tests/validation/CL/GEMMLowp.cpp b/tests/validation/CL/GEMMLowp.cpp
index 8aa81d0962..41a441c3d2 100644
--- a/tests/validation/CL/GEMMLowp.cpp
+++ b/tests/validation/CL/GEMMLowp.cpp
@@ -147,6 +147,65 @@ TEST_SUITE_END() // MatrixMultiplyCore
 
 TEST_SUITE(OutputStage)
 
+TEST_SUITE(QuantizeDownInt32Scale)
+
+TEST_SUITE(QASYMM8)
+
+const auto quantize_down_int32_to_uint8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2,
+                                                      3)
+                                                      * framework::dataset::make("min", 0) * framework::dataset::make("max", 255) * framework::dataset::make("addBias", { false, true });
+
+const auto quantize_down_int32_to_uint8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1,
+                                                           2)
+                                                           * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 173) * framework::dataset::make("addBias", { false, true });
+
+using CLGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage>;
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+TEST_SUITE(BoundedReLu)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+TEST_SUITE_END() // BoundedReLu
+TEST_SUITE_END() // QASYMM8
+
+TEST_SUITE(QASYMM8_SIGNED)
+
+const auto quantize_down_int32_to_int8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2,
+                                                     3)
+                                                     * framework::dataset::make("min", -128) * framework::dataset::make("max", 127) * framework::dataset::make("addBias", { false, true });
+
+const auto quantize_down_int32_to_int8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1,
+                                                          2)
+                                                          * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", -100, -98) * framework::dataset::make("max", 71, 73) * framework::dataset::make("addBias", { false, true });
+
+using CLGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToInt8ScaleValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage>;
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_cases))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+TEST_SUITE(BoundedReLu)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_relu_cases))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+TEST_SUITE_END() // BoundedReLu
+TEST_SUITE_END() // QASYMM8_SIGNED
+TEST_SUITE_END() // QuantizeDownInt32Scale
+
 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)
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index de30bd5451..c3747ddd24 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -165,7 +165,9 @@ TEST_SUITE_END() // MatrixMultiplyCore
 
 TEST_SUITE(OutputStage)
 
-TEST_SUITE(QuantizeDownInt32ToUint8Scale)
+TEST_SUITE(QuantizeDownInt32Scale)
+
+TEST_SUITE(QASYMM8)
 
 const auto quantize_down_int32_to_uint8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2,
                                                       3)
@@ -175,7 +177,7 @@ const auto quantize_down_int32_to_uint8_scale_relu_cases = framework::dataset::m
                                                            2)
                                                            * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 174) * framework::dataset::make("addBias", { false, true });
 
-using NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToUint8Scale>;
+using NEGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpOutputStage>;
 
 // *INDENT-OFF*
 // clang-format off
@@ -198,85 +200,112 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
     framework::dataset::make("Expected", { true, false })),
     a_info, b_info, output_info, min, max, expected)
 {
+
+    GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo();
+    output_stage.type        = GEMMLowpOutputStageType::QUANTIZE_DOWN;
+    output_stage.gemmlowp_min_bound        = min;
+    output_stage.gemmlowp_max_bound        = max;
+    output_stage.output_data_type = DataType::QASYMM8;
+
     // Lock tensors
-    Status status =  NEGEMMLowpQuantizeDownInt32ToUint8Scale::validate(&a_info.clone()->set_is_resizable(false),
+    Status status =  NEGEMMLowpOutputStage::validate(&a_info.clone()->set_is_resizable(false),
                                                                      &b_info.clone()->set_is_resizable(false),
                                                                      &output_info.clone()->set_is_resizable(false),
-                                                                     min,
-                                                                     max);
+                                                                     output_stage);
     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_uint8_scale_cases),
-               shape, result_offset, result_mult_int, result_shift, min, max, add_bias)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases))
 {
-    TensorShape shape_bias(shape[0]);
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
 
-    // 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::QASYMM8);
+TEST_SUITE(BoundedReLu)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
 
-    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);
+TEST_SUITE_END() // BoundedReLu
 
-    // Create and configure function
-    NEGEMMLowpQuantizeDownInt32ToUint8Scale output_stage;
-    output_stage.configure(&in, add_bias ? &bias : nullptr, &out, result_offset, result_mult_int, result_shift, min, max);
+TEST_SUITE_END() // QASYMM8
 
-    // 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);
+TEST_SUITE(QASYMM8_SIGNED)
 
-    // 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);
-    }
+const auto quantize_down_int32_to_int8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2,
+                                                     3)
+                                                     * framework::dataset::make("min", 0) * framework::dataset::make("max", 0) * framework::dataset::make("addBias", { false, true });
 
-    // Validate padding
-    const PaddingSize padding(0);
-    validate(in.info()->padding(), padding);
-    validate(out.info()->padding(), padding);
+const auto quantize_down_int32_to_int8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1,
+                                                          2)
+                                                          * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", -100, -98) * framework::dataset::make("max", 71, 74) * framework::dataset::make("addBias", { false, true });
 
-    if(add_bias)
-    {
-        validate(bias.info()->padding(), padding);
-    }
-}
+using NEGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToInt8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpOutputStage>;
 
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases))
+// *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::QASYMM8_SIGNED),
+                                            TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED),
+                                            TensorInfo(TensorShape(20U, 13U), 1, DataType::S32),
+                                           })),
+    framework::dataset::make("Min",{        -10,
+                                            -200,
+                                            -113,
+                                           })),
+    framework::dataset::make("Max",{        105,
+                                            300,
+                                            -18,
+                                           })),
+    framework::dataset::make("Expected", { true, false, false })),
+    a_info, b_info, output_info, min, max, expected)
 {
-    // Validate output
-    validate(Accessor(_target), _reference);
+    GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo();
+    output_stage.type        = GEMMLowpOutputStageType::QUANTIZE_DOWN;
+    output_stage.gemmlowp_min_bound        = min;
+    output_stage.gemmlowp_max_bound        = max;
+    output_stage.output_data_type = DataType::QASYMM8_SIGNED;
+
+    // Lock tensors
+    Status status =  NEGEMMLowpOutputStage::validate(&a_info.clone()->set_is_resizable(false),
+                                                                     &b_info.clone()->set_is_resizable(false),
+                                                                     &output_info.clone()->set_is_resizable(false),
+                                                                     output_stage);
+    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
 }
+// clang-format on
+// *INDENT-ON*
 
-FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), quantize_down_int32_to_uint8_scale_cases))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_cases))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 
 TEST_SUITE(BoundedReLu)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_relu_cases))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), quantize_down_int32_to_uint8_scale_relu_cases))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
 TEST_SUITE_END() // BoundedReLu
 
-TEST_SUITE_END() // QuantizeDownInt32ToUint8Scale
+TEST_SUITE_END() // QASYMM8_SIGNED
+
+TEST_SUITE_END() // QuantizeDownInt32Scale
 
 TEST_SUITE(QuantizeDownInt32ToUint8ScaleByFixedPoint)
 
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index be9ce96dcb..e3dc7381fc 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -301,8 +301,16 @@ protected:
         TensorType c = create_tensor<TensorType>(shape, DataType::QASYMM8, 1);
 
         // Create and configure function
-        FunctionType output_stage;
-        output_stage.configure(&a, add_bias ? &b : nullptr, &c, result_offset, result_mult_int, result_shift, min, max);
+        FunctionType            output_stage;
+        GEMMLowpOutputStageInfo output_stage_info = GEMMLowpOutputStageInfo();
+        output_stage_info.type                    = GEMMLowpOutputStageType::QUANTIZE_DOWN;
+        output_stage_info.gemmlowp_offset         = result_offset;
+        output_stage_info.gemmlowp_multiplier     = result_mult_int;
+        output_stage_info.gemmlowp_shift          = result_shift;
+        output_stage_info.gemmlowp_min_bound      = min;
+        output_stage_info.gemmlowp_max_bound      = max;
+        output_stage_info.output_data_type        = DataType::QASYMM8;
+        output_stage.configure(&a, add_bias ? &b : nullptr, &c, output_stage_info);
 
         ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -367,6 +375,108 @@ protected:
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType>
+class GEMMLowpQuantizeDownInt32ToInt8ScaleValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias)
+    {
+        _target    = compute_target(shape, result_offset, result_mult_int, result_shift, min, max, add_bias);
+        _reference = compute_reference(shape, result_offset, result_mult_int, 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_offset, int32_t result_mult_int, 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::QASYMM8_SIGNED, 1);
+
+        // Create and configure function
+        FunctionType            output_stage;
+        GEMMLowpOutputStageInfo output_stage_info = GEMMLowpOutputStageInfo();
+        output_stage_info.type                    = GEMMLowpOutputStageType::QUANTIZE_DOWN;
+        output_stage_info.gemmlowp_offset         = result_offset;
+        output_stage_info.gemmlowp_multiplier     = result_mult_int;
+        output_stage_info.gemmlowp_shift          = result_shift;
+        output_stage_info.gemmlowp_min_bound      = min;
+        output_stage_info.gemmlowp_max_bound      = max;
+        output_stage_info.output_data_type        = DataType::QASYMM8_SIGNED;
+        output_stage.configure(&a, add_bias ? &b : nullptr, &c, output_stage_info);
+
+        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<int8_t> compute_reference(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, 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);
+
+        const std::vector<int32_t> result_mult_int_vec = { result_mult_int };
+        const std::vector<int32_t> result_shift_vec    = { result_shift };
+
+        if(add_bias)
+        {
+            // Fill bias
+            fill(b, 1);
+
+            return reference::gemmlowp_quantize_down_scale<int32_t, int8_t>(a, b, result_offset, result_mult_int_vec, result_shift_vec, min, max);
+        }
+        else
+        {
+            return reference::gemmlowp_quantize_down_scale<int32_t, int8_t>(a, result_offset, result_mult_int_vec, result_shift_vec, min, max);
+        }
+    }
+
+    TensorType           _target{};
+    SimpleTensor<int8_t> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType>
 class GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture : public framework::Fixture
 {
 public: