COMPMID-1451: Perform CLOutputStage using floats.

Change-Id: Ic8312a5b6790aa7cd4468d42f08d557ad40e9441
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/154570
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

12 files changed, 474 insertions, 24 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 298cf5241f..1e456fa17e 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -71,6 +71,7 @@
 #include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.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"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h
new file mode 100644
index 0000000000..5a5d3938b7
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h
@@ -0,0 +1,100 @@
+/*
+ * Copyright (c) 2018 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_CLGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFLOATKERNEL_H__
+#define __ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFLOATKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ICLTensor;
+
+/** OpenCL 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 CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Requantize
+ *  -# Add offset to each result
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
+ */
+class CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel : public ICLKernel
+{
+public:
+    /** Constructor */
+    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel(const CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &operator=(const CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel(CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &operator=(CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &&) = 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]  multiplier      Float multiplier to be multiplied to each element of the input matrix
+     * @param[in]  offset          Offset to be applied to result before converting it back to QASYMM8
+     * @param[in]  min             (Optional) Min value used to saturate down the output result before converting back to QASYMM8
+     * @param[in]  max             (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
+     *                             Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     * @param[in]  output_3d_depth (Optional) Depth of output in 3D (Defaults to 1)
+     */
+    void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, float multiplier, int offset,
+                   int min = 0, int max = 0, unsigned int output_3d_depth = 1);
+    /** 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
+     * @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
+     * @param[in] output_3d_depth (Optional)  Depth of output in 3D (Defaults to 1)
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
+                           int min = 0, int max = 0, unsigned int output_3d_depth = 1);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor *_input;
+    const ICLTensor *_bias;
+    ICLTensor       *_output;
+    bool             _reinterpret_as_3d;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFLOATKERNEL_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
index 958e70fca4..48b880174d 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
@@ -178,15 +178,15 @@ private:
     static Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, int gemm_3d_depth = 1, bool skip_im2col = false);
 
 private:
-    CLMemoryGroup                                       _memory_group;
-    CLConvolutionLayerReshapeWeights                    _reshape_weights;
-    CLIm2ColKernel                                      _im2col_kernel;
-    CLGEMM                                              _mm_gemm;
-    CLGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
-    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
-    CLCol2ImKernel                                      _col2im_kernel;
-    CLActivationLayer                                   _activationlayer_function;
-    CLArithmeticAdditionKernel                          _add_bias_kernel;
+    CLMemoryGroup                                  _memory_group;
+    CLConvolutionLayerReshapeWeights               _reshape_weights;
+    CLIm2ColKernel                                 _im2col_kernel;
+    CLGEMM                                         _mm_gemm;
+    CLGEMMLowpMatrixMultiplyCore                   _mm_gemmlowp;
+    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat _gemmlowp_output_stage;
+    CLCol2ImKernel                                 _col2im_kernel;
+    CLActivationLayer                              _activationlayer_function;
+    CLArithmeticAdditionKernel                     _add_bias_kernel;
 
     const ICLTensor *_original_weights;
 
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
index dca00f027e..51fcbe9392 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
@@ -150,5 +150,47 @@ public:
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0, unsigned int output_3d_depth = 1);
 };
+
+/** Basic function to execute CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat on OpenCL.
+ *
+ *  This function calls the following OpenCL kernels:
+ *
+ * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel
+ *
+ * @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 CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat : 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
+     * @param[in]  multiplier      Float multiplier to be multiplied to each element of the input matrix
+     * @param[in]  offset          Offset to be applied to result before converting it back to QASYMM8
+     * @param[in]  min             (Optional) Min value used to saturate down the output result before converting back to QASYMM8
+     * @param[in]  max             (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
+     *                             Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     * @param[in]  output_3d_depth (Optional) Depth of output in 3D (Defaults to 1)
+     */
+    void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, float multiplier, int offset, int min = 0, int max = 0, unsigned int output_3d_depth = 1);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
+     *
+     * @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
+     * @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
+     * @param[in] output_3d_depth (Optional) Depth of output in 3D (Defaults to 1)
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0, unsigned int output_3d_depth = 1);
+};
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H__ */
\ No newline at end of file
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 12a7c38dfd..880963de7b 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -269,6 +269,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemmlowp_offset_contribution", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down_fixedpoint", "gemmlowp.cl" },
+    { "gemmlowp_output_stage_quantize_down_float", "gemmlowp.cl" },
     { "harris_score_3x3", "harris_corners.cl" },
     { "harris_score_5x5", "harris_corners.cl" },
     { "harris_score_7x7", "harris_corners.cl" },
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index 0fc3868341..80b5d00cf2 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -2276,3 +2276,94 @@ __kernel void gemmlowp_output_stage_quantize_down_fixedpoint(TENSOR3D_DECLARATIO
     vstore16(res, 0, dst.ptr);
 }
 #endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
+
+#if defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)
+/** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Requantize
+ *  -# 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.
+ *
+ * @attention The offset and scalar scale factor must be passed at compile time using -DRESULT_OFFSET, -DREAL_MULTIPLIER
+ *
+ * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time
+ * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.
+ *       These values can be used to implement "rectified linear unit" activation functions
+ *
+ * @param[in]  src_ptr                              Pointer to the source tensor. Supported data type: S32
+ * @param[in]  src_stride_x                         Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                           src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                         Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                           src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                         Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                           src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes    The offset of the first element in the source tensor
+ * @param[in]  biases_ptr                           Pointer to the biases tensor. Supported data type: same as @p src_ptr
+ * @param[in]  biases_stride_x                      Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                        biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes The offset of the first element in the biases tensor
+ * @param[out] dst_ptr                              Pointer to the destination tensor Supported data type: QASYMM8
+ * @param[in]  dst_stride_x                         Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  dst_step_x                           dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                         Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                           dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                         Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                           src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_stride_w                         Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  dst_step_w                           src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes    The offset of the first element in the destination tensor
+ */
+__kernel void gemmlowp_output_stage_quantize_down_float(TENSOR3D_DECLARATION(src),
+#if defined(ADD_BIAS)
+                                                        VECTOR_DECLARATION(biases),
+#endif // defined(ADD_BIAS)
+#if defined(DST_HEIGHT)
+                                                        TENSOR4D_DECLARATION(dst))
+#else  // defined(DST_HEIGHT)
+                                                        TENSOR3D_DECLARATION(dst))
+#endif // defined(DST_HEIGHT)
+{
+    // Compute source and destination addresses
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+#if defined(DST_HEIGHT)
+    Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(dst, 1);
+    dst.ptr += get_global_id(0) * dst_step_x + (get_global_id(1) % DST_HEIGHT) * dst_step_y + (get_global_id(1) / DST_HEIGHT) * dst_step_z + get_global_id(2) * dst_step_w;
+#else  // defined(DST_HEIGHT)
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+#endif // defined(DST_HEIGHT)
+
+#if defined(ADD_BIAS)
+    Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
+#endif // defined(ADD_BIAS)
+
+    int16 input_values = vload16(0, (__global int *)src.ptr);
+
+#if defined(ADD_BIAS)
+    // Add bias
+    const int16 biases_values = vload16(0, (__global int *)biases.ptr);
+    input_values += (int16)biases_values;
+#endif // defined(ADD_BIAS)
+
+    // Convert to float
+    float16 input_values_f = convert_float16(input_values);
+    input_values_f         = round(input_values_f * (float)REAL_MULTIPLIER + (float)OUTPUT_OFFSET);
+
+    uchar16 res = convert_uchar16_sat(input_values_f);
+
+#if defined(MIN_BOUND)
+    res = max(res, (uchar16)MIN_BOUND);
+#endif // defined(MIN_BOUND)
+#if defined(MAX_BOUND)
+    res = min(res, (uchar16)MAX_BOUND);
+#endif // defined(MAX_BOUND)
+
+    // Store the result
+    vstore16(res, 0, dst.ptr);
+}
+#endif // defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
new file mode 100644
index 0000000000..f0096bd3ad
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
@@ -0,0 +1,207 @@
+/*
+ * Copyright (c) 2018 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/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
+                          int min, int max, unsigned int output_3d_depth)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
+    ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
+    ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || 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)
+    {
+        const TensorShape output_shape       = arm_compute::misc::shape_calculator::compute_output_stage_shape(*input, output_3d_depth, true);
+        const TensorInfo  tensor_info_output = output->clone()->set_tensor_shape(output_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_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 = 16;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+
+    bool window_changed = update_window_and_padding(win,
+                                                    input_access);
+
+    if(output->total_size() != 0)
+    {
+        Window                 win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+        AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
+        window_changed = window_changed || update_window_and_padding(win_out, output_result_access);
+
+        output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    }
+
+    if(bias != nullptr)
+    {
+        AccessWindowStatic bias_access(bias, 0, 0, ceil_to_multiple(bias->dimension(0), num_elems_processed_per_iteration), bias->tensor_shape()[1]);
+        window_changed = window_changed || update_window_and_padding(win, bias_access);
+    }
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+class Coordinates;
+} // namespace arm_compute
+
+CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel()
+    : _input(nullptr), _bias(nullptr), _output(nullptr), _reinterpret_as_3d(false)
+{
+}
+
+Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
+                                                                      int min, int max, unsigned int output_3d_depth)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max, output_3d_depth));
+    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 CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
+                                                                     float multiplier, int offset,
+                                                                     int min, int max, unsigned int output_3d_depth)
+{
+    // Perform validate step
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Output auto inizialitation if not yet initialized
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_output_stage_shape(*input->info(), output_3d_depth, true);
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(DataType::QASYMM8).set_tensor_shape(output_shape));
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(),
+                                                  min, max, output_3d_depth));
+
+    _input             = input;
+    _bias              = bias;
+    _output            = output;
+    _reinterpret_as_3d = output_3d_depth > 1;
+
+    // Set the arguments to pass at compile time
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DREAL_MULTIPLIER=" + float_to_string_with_full_precision(multiplier));
+    build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(offset));
+    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");
+    build_opts.add_option_if(_reinterpret_as_3d, "-DDST_HEIGHT=" + support::cpp11::to_string(input->info()->tensor_shape().y() / output_3d_depth));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_output_stage_quantize_down_float", 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 CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::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);
+    }
+
+    if(_reinterpret_as_3d)
+    {
+        // Create output window
+        Window window_out;
+        window_out.use_tensor_dimensions(_output->info()->tensor_shape());
+        Window collapsed_out = window_out.collapse_if_possible(window_out, 3);
+        Window slice_out     = collapsed.first_slice_window_4D();
+
+        do
+        {
+            unsigned int idx = 0;
+            add_3D_tensor_argument(idx, _input, slice);
+            add_4D_tensor_argument(idx1, _output, slice_out);
+            enqueue(queue, *this, slice);
+        }
+        while(collapsed.slide_window_slice_3D(slice) && collapsed_out.slide_window_slice_4D(slice_out));
+    }
+    else
+    {
+        do
+        {
+            unsigned int idx = 0;
+            add_3D_tensor_argument(idx, _input, slice);
+            add_3D_tensor_argument(idx1, _output, slice);
+            enqueue(queue, *this, slice);
+        }
+        while(collapsed.slide_window_slice_3D(slice));
+    }
+}
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index f41a12ae48..61180fd5d3 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -284,17 +284,14 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
     {
         const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
 
-        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
-        int   output_multiplier, output_shift;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
         if(!_skip_col2im)
         {
             _memory_group.manage(&_tmp_output);
             gemm_output_staged_to_use = &_tmp_output;
         }
 
-        _gemmlowp_output_stage.configure(gemm_output_to_use, biases, gemm_output_staged_to_use, output_multiplier, output_shift, output_quant_info.offset);
+        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
+        _gemmlowp_output_stage.configure(gemm_output_to_use, biases, gemm_output_staged_to_use, multiplier, output_quant_info.offset);
     }
 
     if(!_skip_col2im)
@@ -448,17 +445,12 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
 
     if(is_quantized)
     {
-        float multiplier = input->quantization_info().scale * weights_to_use->quantization_info().scale / output->quantization_info().scale;
-        int   output_multiplier, output_shift;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
         if(!skip_col2im)
         {
             tmp_info = TensorInfo(gemm_output_to_use->tensor_shape(), 1, DataType::QASYMM8);
             tmp_info.set_quantization_info(output->quantization_info());
             gemm_output_staged_to_use = &tmp_info;
         }
-
         // Validate output stage for quantized case
         CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(gemm_output_to_use, biases, gemm_output_staged_to_use);
     }
diff --git a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
index b18d23fac9..f5dc655776 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/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
 #include "support/ToolchainSupport.h"
 
@@ -56,4 +57,19 @@ Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(const ITens
 {
     return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, min, max, output_3d_depth);
 }
+
+void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
+                                                               float multiplier, int offset,
+                                                               int min, int max, unsigned int output_3d_depth)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel>();
+    k->configure(input, bias, output, multiplier, offset, min, max, output_3d_depth);
+    _kernel = std::move(k);
+}
+
+Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
+                                                                int min, int max, unsigned int output_3d_depth)
+{
+    return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::validate(input, bias, output, min, max, output_3d_depth);
+}
 } // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/CL/BatchNormalizationLayer.cpp b/tests/validation/CL/BatchNormalizationLayer.cpp
index cbf3c7092d..5aef357b4d 100644
--- a/tests/validation/CL/BatchNormalizationLayer.cpp
+++ b/tests/validation/CL/BatchNormalizationLayer.cpp
@@ -48,9 +48,9 @@ namespace validation
 {
 namespace
 {
-RelativeTolerance<float>           rel_tolerance_f32(0.05f);    /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
-constexpr AbsoluteTolerance<float> abs_tolerance_f32(0.00001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
-constexpr AbsoluteTolerance<float> tolerance_f16(0.01f);        /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
+RelativeTolerance<float>           rel_tolerance_f32(0.05f);   /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+constexpr AbsoluteTolerance<float> abs_tolerance_f32(0.0001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+constexpr AbsoluteTolerance<float> tolerance_f16(0.01f);       /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
 const auto                         act_infos = framework::dataset::make("ActivationInfo",
 {
     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp
index 5c96cd4c59..0274bed977 100644
--- a/tests/validation/CL/ConvolutionLayer.cpp
+++ b/tests/validation/CL/ConvolutionLayer.cpp
@@ -48,7 +48,7 @@ namespace
 constexpr AbsoluteTolerance<float>  absolute_tolerance_float(0.0001f);    /**< Absolute Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
 RelativeTolerance<float>            tolerance_f32(0.05f);                 /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
 RelativeTolerance<half_float::half> tolerance_f16(half_float::half(0.2)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
-constexpr AbsoluteTolerance<float>  tolerance_qasymm8(0.0);               /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
+constexpr AbsoluteTolerance<float>  tolerance_qasymm8(1);                 /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
 constexpr float                     tolerance_num = 0.07f;                /**< Tolerance number */
 
 /** CNN data types */
diff --git a/tests/validation/CL/DilatedConvolutionLayer.cpp b/tests/validation/CL/DilatedConvolutionLayer.cpp
index 9ebde38bcf..2ddc031e14 100644
--- a/tests/validation/CL/DilatedConvolutionLayer.cpp
+++ b/tests/validation/CL/DilatedConvolutionLayer.cpp
@@ -45,7 +45,7 @@ namespace
 {
 RelativeTolerance<float>            rel_tolerance_f32(0.05f);                 /**< Relative tolerance value for comparing reference's output against implementation's output for DataType::F32 */
 RelativeTolerance<half_float::half> rel_tolerance_f16(half_float::half(0.2)); /**< Relative tolerance value for comparing reference's output against implementation's output for DataType::F16 */
-constexpr AbsoluteTolerance<float>  abs_tolerance_qasymm8(0.0);               /**< Relative tolerance value for comparing reference's output against implementation's output for quantized data types */
+constexpr AbsoluteTolerance<float>  abs_tolerance_qasymm8(1);                 /**< Relative tolerance value for comparing reference's output against implementation's output for quantized data types */
 constexpr float                     abs_tolerance_f32 = 0.001f;               /**< Absolute tolerance value for comparing reference's output against implementation's output for DataType::F32 */
 constexpr float                     abs_tolerance_f16 = 0.3f;                 /**< Absolute tolerance value for comparing reference's output against implementation's output for DataType::F16 */
 constexpr float                     tolerance_num_f16 = 0.07f;                /**< Tolerance number for FP16 */