COMPMID-2306: CLDepthwiseConvolution: support for QUANT8_PER_CHANNEL_SYMM

Change-Id: I18c886400daa2dcba0b91011bc4e503d807a4732
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2143
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp b/src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp
index f232f6c..e883e8f 100644
--- a/src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp
+++ b/src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp
@@ -113,21 +113,7 @@
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size));
     build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(input->info()->dimension(2)));
     build_opts.add_option("-DLAST_ACCESSED=" + support::cpp11::to_string(std::max(static_cast<int>(channels - vec_size), 0)));
-
-    switch(input->info()->element_size())
-    {
-        case 1:
-            build_opts.add_option("-DDATA_TYPE=uchar");
-            break;
-        case 2:
-            build_opts.add_option("-DDATA_TYPE=ushort");
-            break;
-        case 4:
-            build_opts.add_option("-DDATA_TYPE=uint");
-            break;
-        default:
-            ARM_COMPUTE_ERROR("Data type not supported");
-    }
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
 
     // Create kernel
     std::string kernel_name = "channel_shuffle_" + lower_string(string_from_data_layout(data_layout));
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
index 42e5fbc..a2f4a91 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
@@ -37,13 +37,15 @@
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 
-using namespace arm_compute;
+namespace arm_compute
+{
 using namespace arm_compute::misc::shape_calculator;
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
-                          const ActivationLayerInfo &act_info, const Size2D dilation)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                          const PadStrideInfo &conv_info, unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D dilation,
+                          const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
@@ -52,7 +54,6 @@
                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU)
                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LOGISTIC),
                                     "For QASYMM8 only logistic, relu, lower bounded relu and lower-upper bounded relu are supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != 3 || weights->dimension(1) != 3);
     ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1 || conv_info.stride().first > 3);
 
@@ -74,28 +75,43 @@
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
     }
 
+    if(is_qasymm)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output_multipliers, output_shifts);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32);
+        ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
+
+        if(is_data_type_quantized_per_channel(weights->data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QSYMM8_PER_CHANNEL);
+            ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(2) != output_multipliers->dimension(0));
+            ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(2) != output_shifts->dimension(0));
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+            ARM_COMPUTE_RETURN_ERROR_ON(1 != output_multipliers->dimension(0));
+            ARM_COMPUTE_RETURN_ERROR_ON(1 != output_shifts->dimension(0));
+        }
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+    }
+
     if(output->total_size() != 0)
     {
         const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
     }
 
-    if(is_qasymm)
-    {
-        const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = (output->total_size() != 0) ? output->quantization_info().uniform() : iq_info;
-
-        float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
-        ARM_COMPUTE_UNUSED(multiplier);
-        ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
-    }
-
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
-                                                        GPUTarget gpu_target, std::string &kernel_name, const Size2D dilation)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info,
+                                                        unsigned int depth_multiplier, GPUTarget gpu_target, std::string &kernel_name, const Size2D dilation)
 {
     // Output auto inizialitation if not yet initialized
     const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
@@ -182,9 +198,9 @@
     }
     else
     {
-        const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device());
+        const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device()) && !is_data_type_quantized_per_channel(weights->data_type());
 
-        kernel_name = is_qasymm ? "dwc_3x3_native_qasymm8" : "depthwise_convolution_3x3";
+        kernel_name = is_qasymm ? "dwc_3x3_native_quantized8" : "depthwise_convolution_3x3";
         kernel_name += (is_qasymm && is_dot8_supported ? "_dot8" : "");
         kernel_name += (is_qasymm ? "_nchw" : "");
 
@@ -224,23 +240,28 @@
     return _border_size;
 }
 
-void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                                                         unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
+void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
+                                                         const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation,
+                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info, dilation));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(),
+                                                  conv_info, depth_multiplier, act_info, dilation,
+                                                  (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
+                                                  (output_shifts != nullptr) ? output_shifts->info() : nullptr));
 
-    bool is_qasymm = is_data_type_quantized_asymmetric(input->info()->data_type());
-
-    _input         = input;
-    _output        = output;
-    _weights       = weights;
-    _biases        = biases;
-    _conv_stride_x = conv_info.stride().first;
-    _conv_stride_y = conv_info.stride().second;
-    _conv_pad_left = conv_info.pad_left();
-    _conv_pad_top  = conv_info.pad_top();
-    _border_size   = BorderSize(_conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), _conv_pad_left);
+    _input              = input;
+    _output             = output;
+    _weights            = weights;
+    _biases             = biases;
+    _conv_stride_x      = conv_info.stride().first;
+    _conv_stride_y      = conv_info.stride().second;
+    _conv_pad_left      = conv_info.pad_left();
+    _conv_pad_top       = conv_info.pad_top();
+    _border_size        = BorderSize(_conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), _conv_pad_left);
+    _output_multipliers = output_multipliers;
+    _output_shifts      = output_shifts;
+    _is_quantized       = is_data_type_quantized_asymmetric(input->info()->data_type());
 
     // Configure kernel window
     std::string     kernel_name;
@@ -260,24 +281,21 @@
     build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
 
-    if(is_qasymm)
+    if(_is_quantized)
     {
         const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
         const UniformQuantizationInfo wq_info = _weights->info()->quantization_info().uniform();
         const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
 
-        float multiplier        = iq_info.scale * wq_info.scale / oq_info.scale;
-        int   output_multiplier = 0;
-        int   output_shift      = 0;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
+        const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->info()->data_type());
+        const bool is_dot8_supported        = dot8_supported(CLKernelLibrary::get().get_device()) && !is_quantized_per_channel;
         build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(_conv_stride_y));
         build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-iq_info.offset));
         build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-wq_info.offset));
         build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(oq_info.offset));
         build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * iq_info.offset * wq_info.offset));
-        build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
-        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
+        build_opts.add_option_if(is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION");
+        build_opts.add_option_if(is_dot8_supported, "-DIS_DOT8");
 
         if(act_info.enabled())
         {
@@ -293,6 +311,10 @@
             build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
             build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
         }
+
+        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+        build_opts.add_option("-DWEIGHTS_TYPE=" + get_cl_type_from_data_type(weights->info()->data_type()));
+        build_opts.add_option("-DWEIGHTS_PROMOTED_TYPE=" + get_cl_promoted_type_from_data_type(weights->info()->data_type()));
     }
     else
     {
@@ -323,12 +345,15 @@
     _config_id += support::cpp11::to_string(output->info()->dimension(1));
 }
 
-Status CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                                          unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target, const Size2D &dilation)
+Status CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                                          const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target,
+                                                          const Size2D &dilation, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
     std::string kernel_name;
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(), conv_info, depth_multiplier, gpu_target, kernel_name, dilation).first);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation, output_multipliers, output_shifts));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(),
+                                                              conv_info, depth_multiplier, gpu_target, kernel_name, dilation)
+                                .first);
 
     return Status{};
 }
@@ -353,18 +378,28 @@
     slice_weights.set_dimension_step(Window::DimX, 0);
     slice_weights.set_dimension_step(Window::DimY, 0);
 
+    unsigned int idx = 3 * num_arguments_per_3D_tensor();
+
+    // Set output multipliers in case of quantized data type
+    if(_is_quantized)
+    {
+        Window slice;
+        slice.use_tensor_dimensions(_output_multipliers->info()->tensor_shape());
+        add_1D_tensor_argument(idx, _output_multipliers, slice);
+        add_1D_tensor_argument(idx, _output_shifts, slice);
+    }
+
     // Set biases
     if(_biases != nullptr)
     {
-        unsigned int idx = 3 * num_arguments_per_3D_tensor();
-        Window       slice_biases;
+        Window slice_biases;
         slice_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
         add_1D_tensor_argument(idx, _biases, slice_biases);
     }
 
     do
     {
-        unsigned int idx = 0;
+        idx = 0;
         add_3D_tensor_argument(idx, _input, slice_in);
         add_3D_tensor_argument(idx, _output, slice_out);
         add_3D_tensor_argument(idx, _weights, slice_weights);
@@ -373,3 +408,4 @@
     }
     while(collapsed.slide_window_slice_3D(slice_out) && collapsed_in.slide_window_slice_3D(slice_in));
 }
+} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index b8b144d..d5f37f3 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -41,17 +41,18 @@
 {
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
-                          const ActivationLayerInfo &act_info, const Size2D &dilation)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                          const PadStrideInfo &conv_info, unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation,
+                          const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::QASYMM8);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.enabled()) && (input->data_type() == DataType::QASYMM8) && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.enabled()) && (input->data_type() == DataType::QASYMM8)
+                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU)
                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LOGISTIC),
                                     "For QASYMM8 only logistic, relu, lower bounded relu and lower-upper bounded relu are supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1); // COMPMID-1071 Add depth multiplier support for NHWC
 
     ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
@@ -63,26 +64,47 @@
     const size_t weights_width  = 3;
     const size_t weights_height = 3;
 
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(
+                                         *input, TensorInfo(TensorShape(weights_width, weights_height), 1, weights->data_type()).set_data_layout(DataLayout::NCHW), conv_info, depth_multiplier, dilation);
     if(is_qasymm)
     {
         DepthwiseConvolutionReshapeInfo info;
         info.c0 = 4;
         ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(0) / info.c0) != weights_width * weights_height);
+
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output_multipliers, output_shifts);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32);
+        ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
+
+        if(is_data_type_quantized_per_channel(weights->data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON(output_shape[0] != output_multipliers->dimension(0));
+            ARM_COMPUTE_RETURN_ERROR_ON(output_shape[0] != output_shifts->dimension(0));
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+            ARM_COMPUTE_RETURN_ERROR_ON(1 != output_multipliers->dimension(0));
+            ARM_COMPUTE_RETURN_ERROR_ON(1 != output_shifts->dimension(0));
+        }
     }
     else
     {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
         ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(1) != weights_width) || (weights->dimension(2) != weights_height));
     }
 
     if(biases != nullptr)
     {
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != output_shape[0]);
         if(is_qasymm)
         {
             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
         }
         else
         {
-            ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
         }
 
@@ -91,27 +113,15 @@
 
     if(output->total_size() != 0)
     {
-        const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(
-                                             *input, TensorInfo(TensorShape(weights_width, weights_height), 1, weights->data_type()).set_data_layout(DataLayout::NCHW), conv_info, depth_multiplier, dilation);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
     }
 
-    if(is_qasymm)
-    {
-        const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = (output->total_size() != 0) ? output->quantization_info().uniform() : iq_info;
-
-        float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
-        ARM_COMPUTE_UNUSED(multiplier);
-        ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
-    }
-
     return Status{};
 }
 
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *output,
-                                                        const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
+                                                        const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
+                                                        ITensorInfo *output_multipliers, ITensorInfo *output_shifts)
 {
     const size_t weights_width  = 3;
     const size_t weights_height = 3;
@@ -144,7 +154,17 @@
 
     if(is_qasymm)
     {
-        window_changed = update_window_and_padding(win, input_access, output_access);
+        if((output_multipliers != nullptr) && (output_shifts != nullptr))
+        {
+            AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_accessed_per_iteration);
+            AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_accessed_per_iteration);
+            window_changed = window_changed || update_window_and_padding(win, input_access, output_access, output_multipliers_access, output_shifts_access);
+        }
+        else
+        {
+            Status err = ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "output_multipliers and output_shifts must be non-nullptr for quantized input");
+            return std::make_pair(err, win);
+        }
     }
     else
     {
@@ -157,7 +177,6 @@
         AccessWindowHorizontal bias_access(bias, 0, num_elems_accessed_per_iteration);
         window_changed = window_changed || update_window_and_padding(win, bias_access);
     }
-
     output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
 
     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
@@ -175,19 +194,26 @@
     return _border_size;
 }
 
-void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                                                         unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
+void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
+                                                         const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation,
+                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info, dilation));
-    auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, dilation);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(),
+                                                  conv_info, depth_multiplier, act_info, dilation,
+                                                  (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
+                                                  (output_shifts != nullptr) ? output_shifts->info() : nullptr));
+    auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
+                                                    conv_info, depth_multiplier, dilation,
+                                                    (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
+                                                    (output_shifts != nullptr) ? output_shifts->info() : nullptr);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
 
-    const bool is_qasymm              = is_data_type_quantized_asymmetric(input->info()->data_type());
     const bool is_stride_1            = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
     const bool is_stride_1_dilation_1 = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
 
-    const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device());
+    const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->info()->data_type());
+    const bool is_dot8_supported        = dot8_supported(CLKernelLibrary::get().get_device()) && !is_quantized_per_channel;
 
     _input                              = input;
     _output                             = output;
@@ -196,16 +222,19 @@
     _conv_stride_y                      = conv_info.stride().second;
     _num_rows_processed_per_iteration   = is_stride_1_dilation_1 ? 2 : 1;
     _num_planes_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
+    _output_multipliers                 = output_multipliers;
+    _output_shifts                      = output_shifts;
+    _is_quantized                       = is_data_type_quantized_asymmetric(input->info()->data_type());
 
     // If QASYMM8 and the 8 bit dot product is available, force _num_planes_processed_per_iteration to 1
-    if(is_dot8_supported && is_qasymm)
+    if(is_dot8_supported && _is_quantized)
     {
         _num_planes_processed_per_iteration = 1;
     }
 
-    _border_size = BorderSize(is_qasymm && is_stride_1 ? 0 : conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
+    _border_size = BorderSize(_is_quantized && is_stride_1 ? 0 : conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
 
-    const unsigned int num_elems_accessed_per_iteration = is_qasymm ? 4 : (8 / input->info()->element_size());
+    const unsigned int num_elems_accessed_per_iteration = _is_quantized ? 4 : (8 / input->info()->element_size());
 
     CLBuildOptions build_opts;
     build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation())));
@@ -217,24 +246,19 @@
     build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
     build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
 
-    if(is_qasymm)
+    if(_is_quantized)
     {
         const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
         const UniformQuantizationInfo wq_info = _weights->info()->quantization_info().uniform();
         const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
 
-        float multiplier        = iq_info.scale * wq_info.scale / oq_info.scale;
-        int   output_multiplier = 0;
-        int   output_shift      = 0;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
         build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
         build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-iq_info.offset));
         build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-wq_info.offset));
         build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(oq_info.offset));
         build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * iq_info.offset * wq_info.offset));
-        build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
-        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
+        build_opts.add_option_if(is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION");
+        build_opts.add_option_if(is_dot8_supported, "-DIS_DOT8");
 
         if(act_info.enabled())
         {
@@ -250,6 +274,10 @@
             build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
             build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
         }
+
+        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+        build_opts.add_option("-DWEIGHTS_TYPE=" + get_cl_type_from_data_type(weights->info()->data_type()));
+        build_opts.add_option("-DWEIGHTS_PROMOTED_TYPE=" + get_cl_promoted_type_from_data_type(weights->info()->data_type()));
     }
     else
     {
@@ -274,9 +302,9 @@
 
     std::string kernel_name;
     // Create kernel
-    if(is_qasymm)
+    if(_is_quantized)
     {
-        kernel_name = std::string("dwc_3x3_reshaped_qasymm8");
+        kernel_name = std::string("dwc_3x3_reshaped_quantized8");
         kernel_name += (is_dot8_supported && is_stride_1_dilation_1 ? "_dot8" : "");
         kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
         kernel_name += "_nhwc";
@@ -309,13 +337,16 @@
     _config_id += string_from_data_type(input->info()->data_type());
 }
 
-Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                                          unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
+Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                                          const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation,
+                                                          const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation, output_multipliers, output_shifts));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(),
                                                               biases != nullptr ? biases->clone().get() : nullptr,
-                                                              output->clone().get(), conv_info, depth_multiplier, dilation)
+                                                              output->clone().get(), conv_info, depth_multiplier, dilation,
+                                                              (output_multipliers != nullptr) ? output_multipliers->clone().get() : nullptr,
+                                                              (output_shifts != nullptr) ? output_shifts->clone().get() : nullptr)
                                 .first);
 
     return Status{};
@@ -329,7 +360,6 @@
     // Collapse window
     Window       window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
     const size_t total_batches    = _input->info()->tensor_shape().total_size_upper(3);
-    const bool   is_qasymm        = is_data_type_quantized_asymmetric(_input->info()->data_type());
 
     Window win = window_collapsed;
     win.set(Window::DimZ, Window::Dimension(0, std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)) * total_batches, 1));
@@ -344,7 +374,16 @@
     Window slice_in  = win_in.first_slice_window_4D();
     Window slice_out = win.first_slice_window_4D();
 
-    unsigned int idx = 2 * num_arguments_per_4D_tensor() + (is_qasymm ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
+    unsigned int idx = 2 * num_arguments_per_4D_tensor() + (_is_quantized ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
+
+    if(_is_quantized)
+    {
+        Window slice;
+        slice.use_tensor_dimensions(_output_multipliers->info()->tensor_shape());
+        slice.set_dimension_step(Window::DimX, window.x().step());
+        add_1D_tensor_argument(idx, _output_multipliers, slice);
+        add_1D_tensor_argument(idx, _output_shifts, slice);
+    }
 
     if(_biases != nullptr)
     {
@@ -398,7 +437,7 @@
         unsigned int idx = 0;
         add_4D_tensor_argument(idx, _input, slice_in);
         add_4D_tensor_argument(idx, _output, slice_out);
-        if(is_qasymm)
+        if(_is_quantized)
         {
             add_2D_tensor_argument(idx, _weights, slice_out);
         }
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
index 2115fc6..3fc236e 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
@@ -42,13 +42,13 @@
 namespace
 {
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCWeightsKernelInfo &dwc_weights_info,
-                          const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
+                          const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
+                          const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
     ARM_COMPUTE_UNUSED(dwc_info);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1 && dwc_weights_info.n0 != 1);
     ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
     ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().second < 1);
@@ -57,24 +57,53 @@
     ARM_COMPUTE_UNUSED(idx_c);
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_c) != (input->dimension(idx_c) * depth_multiplier));
 
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
+
+    const bool is_quantized = is_data_type_quantized(input->data_type());
+
     if(biases != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != output_shape[idx_c]);
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
 
-        if(is_data_type_quantized(input->data_type()))
+        if(is_quantized)
         {
             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
         }
         else
         {
-            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
         }
     }
 
+    if(is_quantized)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output_multipliers, output_shifts);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32);
+        ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
+
+        if(is_data_type_quantized_per_channel(weights->data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QSYMM8_PER_CHANNEL);
+            ARM_COMPUTE_RETURN_ERROR_ON(output_shape[idx_c] != output_multipliers->dimension(0));
+            ARM_COMPUTE_RETURN_ERROR_ON(output_shape[idx_c] != output_shifts->dimension(0));
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+            ARM_COMPUTE_RETURN_ERROR_ON(1 != output_multipliers->dimension(0));
+            ARM_COMPUTE_RETURN_ERROR_ON(1 != output_shifts->dimension(0));
+        }
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+    }
+
     if(output->total_size() != 0)
     {
-        const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
     }
 
@@ -82,7 +111,8 @@
 }
 
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *output, const DWCWeightsKernelInfo &dwc_weights_info,
-                                                        const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
+                                                        const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
+                                                        ITensorInfo *output_multipliers, ITensorInfo *output_shifts)
 {
     ARM_COMPUTE_UNUSED(dwc_info);
 
@@ -113,6 +143,21 @@
         window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
     }
 
+    if(is_data_type_quantized(input->data_type()))
+    {
+        if((output_multipliers != nullptr) && (output_shifts != nullptr))
+        {
+            AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, n0);
+            AccessWindowHorizontal output_shifts_access(output_shifts, 0, n0);
+            window_changed = window_changed || update_window_and_padding(win, output_multipliers_access, output_shifts_access);
+        }
+        else
+        {
+            Status err = ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "output_multipliers and output_shifts must be non-nullptr for quantized input");
+            return std::make_pair(err, win);
+        }
+    }
+
     output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
 
     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
@@ -121,32 +166,44 @@
 } // namespace
 
 CLDepthwiseConvolutionLayerNativeKernel::CLDepthwiseConvolutionLayerNativeKernel()
-    : _input(nullptr), _weights(nullptr), _biases(nullptr), _output(nullptr), _depth_multiplier(1)
+    : _input(nullptr),
+      _weights(nullptr),
+      _biases(nullptr),
+      _output(nullptr),
+      _depth_multiplier(1),
+      _output_multipliers(nullptr),
+      _output_shifts(nullptr),
+      _is_quantized(false)
 {
 }
 
 void CLDepthwiseConvolutionLayerNativeKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
-                                                        const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
+                                                        const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
+                                                        const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), dwc_weights_info, dwc_info, conv_info, depth_multiplier,
-                                                  dilation));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(),
+                                                  dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation,
+                                                  (output_multipliers != nullptr) ? output_multipliers->info() : nullptr, (output_shifts != nullptr) ? output_shifts->info() : nullptr));
 
-    auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), dwc_weights_info, dwc_info, conv_info, depth_multiplier,
-                                                    dilation);
+    auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
+                                                    dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation,
+                                                    (output_multipliers != nullptr) ? output_multipliers->info() : nullptr, (output_shifts != nullptr) ? output_shifts->info() : nullptr);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
 
-    _input            = input;
-    _output           = output;
-    _weights          = weights;
-    _biases           = biases;
-    _depth_multiplier = depth_multiplier;
+    _input              = input;
+    _output             = output;
+    _weights            = weights;
+    _biases             = biases;
+    _depth_multiplier   = depth_multiplier;
+    _output_multipliers = output_multipliers;
+    _output_shifts      = output_shifts;
+    _is_quantized       = is_data_type_quantized(input->info()->data_type());
 
     const size_t idx_w          = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
     const size_t idx_h          = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
     const size_t weights_width  = weights->info()->dimension(idx_w);
     const size_t weights_height = weights->info()->dimension(idx_h);
-    const bool   is_quantized   = is_data_type_quantized(input->info()->data_type());
 
     CLBuildOptions build_opts;
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
@@ -166,24 +223,18 @@
     build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
     build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
 
-    std::string kernel_name = (is_quantized) ? "dwc_MxN_native_quantized8_nhwc" : "dwc_MxN_native_fp_nhwc";
+    std::string kernel_name = (_is_quantized) ? "dwc_MxN_native_quantized8_nhwc" : "dwc_MxN_native_fp_nhwc";
 
-    if(is_quantized)
+    if(_is_quantized)
     {
         const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
         const UniformQuantizationInfo wq_info = _weights->info()->quantization_info().uniform();
         const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
 
-        float multiplier        = iq_info.scale * wq_info.scale / oq_info.scale;
-        int   output_multiplier = 0;
-        int   output_shift      = 0;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
         build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-iq_info.offset));
         build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-wq_info.offset));
         build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(oq_info.offset));
-        build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
-        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
+        build_opts.add_option_if(is_data_type_quantized_per_channel(weights->info()->data_type()), "-DPER_CHANNEL_QUANTIZATION");
 
         if(dwc_info.activation_info.enabled())
         {
@@ -199,6 +250,9 @@
             build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
             build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
         }
+
+        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+        build_opts.add_option("-DWEIGHTS_TYPE=" + get_cl_type_from_data_type(weights->info()->data_type()));
     }
     else
     {
@@ -228,12 +282,15 @@
 }
 
 Status CLDepthwiseConvolutionLayerNativeKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
-                                                         const DWCWeightsKernelInfo &dwc_weights_info, const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
+                                                         const DWCWeightsKernelInfo &dwc_weights_info, const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info,
+                                                         unsigned int depth_multiplier, const Size2D &dilation, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation, output_multipliers, output_shifts));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(),
                                                               biases != nullptr ? biases->clone().get() : nullptr,
-                                                              output->clone().get(), dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation)
+                                                              output->clone().get(), dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation,
+                                                              output_multipliers != nullptr ? output_multipliers->clone().get() : nullptr,
+                                                              output_shifts != nullptr ? output_shifts->clone().get() : nullptr)
                                 .first);
 
     return Status{};
@@ -255,15 +312,23 @@
         slice_out.set(Window::DimX, Window::Dimension(0, _input->info()->tensor_shape()[0], 1));
     }
 
+    unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_3D_tensor();
+
+    // Set output multipliers in case of quantized data type
+    if(_is_quantized)
+    {
+        add_1D_tensor_argument(idx, _output_multipliers, slice_in);
+        add_1D_tensor_argument(idx, _output_shifts, slice_in);
+    }
+
     if(_biases != nullptr)
     {
-        unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_3D_tensor();
         add_1D_tensor_argument(idx, _biases, slice_in);
     }
 
     do
     {
-        unsigned int idx = 0;
+        idx = 0;
         add_4D_tensor_argument(idx, _input, slice_in);
         add_4D_tensor_argument(idx, _output, slice_out);
         add_3D_tensor_argument(idx, _weights, slice_out);
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp
index 1fd6312..ec889ec 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp
@@ -47,7 +47,6 @@
     const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
 
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
     ARM_COMPUTE_RETURN_ERROR_ON(info.c0 != 4);
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_h) != 3);
@@ -98,10 +97,10 @@
 
     // Build the kernel
     CLBuildOptions build_opts;
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(info.c0));
     build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(0)));
     build_opts.add_option_if(info.transpose, "-DTRANSPOSE");
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
 
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_convolution_reshape_weights", build_opts.options()));
 }
diff --git a/src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.cpp b/src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.cpp
index 72f2ca4..7010dff 100644
--- a/src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.cpp
@@ -37,7 +37,8 @@
 #include "arm_compute/core/Window.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 
-using namespace arm_compute;
+namespace arm_compute
+{
 using namespace arm_compute::misc::shape_calculator;
 
 namespace
@@ -139,21 +140,7 @@
     build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
     build_opts.add_option_if(_reinterpret_input_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(input->info()->dimension(1)));
     build_opts.add_option_if(_reinterpret_input_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(input->info()->dimension(2)));
-
-    switch(input->info()->element_size())
-    {
-        case 1:
-            build_opts.add_option("-DDATA_TYPE=uchar");
-            break;
-        case 2:
-            build_opts.add_option("-DDATA_TYPE=ushort");
-            break;
-        case 4:
-            build_opts.add_option("-DDATA_TYPE=uint");
-            break;
-        default:
-            ARM_COMPUTE_ERROR("Data type not supported");
-    }
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
 
     std::string kernel_name("gemm_reshape_lhs_matrix_");
     kernel_name += lhs_info.transpose ? "t" : "nt";
@@ -219,4 +206,5 @@
         enqueue(queue, *this, slice, lws_hint());
     }
     while(window.slide_window_slice_3D(slice));
-}
\ No newline at end of file
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp b/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp
index 2ca4132..6f6019d 100644
--- a/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp
@@ -37,7 +37,8 @@
 #include "arm_compute/core/Window.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 
-using namespace arm_compute;
+namespace arm_compute
+{
 using namespace arm_compute::misc::shape_calculator;
 
 namespace
@@ -118,21 +119,7 @@
     build_opts.add_option_if(rhs_info.transpose, "-DTRANSPOSE");
     build_opts.add_option_if(rhs_info.interleave, "-DINTERLEAVE");
     build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
-
-    switch(input->info()->element_size())
-    {
-        case 1:
-            build_opts.add_option("-DDATA_TYPE=uchar");
-            break;
-        case 2:
-            build_opts.add_option("-DDATA_TYPE=ushort");
-            break;
-        case 4:
-            build_opts.add_option("-DDATA_TYPE=uint");
-            break;
-        default:
-            ARM_COMPUTE_ERROR("Data type not supported");
-    }
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
 
     std::string kernel_name("gemm_reshape_rhs_matrix_");
     kernel_name += rhs_info.transpose ? "t" : "nt";
@@ -169,4 +156,5 @@
         enqueue(queue, *this, slice, lws_hint());
     }
     while(window.slide_window_slice_3D(slice));
-}
\ No newline at end of file
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLHeightConcatenateLayerKernel.cpp b/src/core/CL/kernels/CLHeightConcatenateLayerKernel.cpp
index ea292c0..85917d3 100644
--- a/src/core/CL/kernels/CLHeightConcatenateLayerKernel.cpp
+++ b/src/core/CL/kernels/CLHeightConcatenateLayerKernel.cpp
@@ -40,7 +40,8 @@
 
 #include <map>
 
-using namespace arm_compute;
+namespace arm_compute
+{
 namespace
 {
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int height_offset, ITensorInfo *output, unsigned int &num_elems_processed_per_iteration)
@@ -102,31 +103,7 @@
 
     // Add build options
     CLBuildOptions build_opts;
-
-    switch(input->info()->element_size())
-    {
-        case 1:
-        {
-            build_opts.add_option("-DDATA_TYPE=uchar");
-            break;
-        }
-        case 2:
-        {
-            build_opts.add_option("-DDATA_TYPE=short");
-            break;
-        }
-        case 4:
-        {
-            build_opts.add_option("-DDATA_TYPE=int");
-            break;
-        }
-        default:
-        {
-            ARM_COMPUTE_ERROR("Unsupported input data type.");
-            break;
-        }
-    }
-
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(_num_elems_processed_per_iteration));
     build_opts.add_option("-DHEIGHT_OFFSET=" + support::cpp11::to_string(_height_offset));
     build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
@@ -164,3 +141,4 @@
     add_4D_tensor_argument(idx, _output, window);
     enqueue(queue, *this, window, lws_hint());
 }
+} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLPermuteKernel.cpp b/src/core/CL/kernels/CLPermuteKernel.cpp
index 9cb72b3..81a810f 100644
--- a/src/core/CL/kernels/CLPermuteKernel.cpp
+++ b/src/core/CL/kernels/CLPermuteKernel.cpp
@@ -52,11 +52,6 @@
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8,
-                                                         DataType::U16, DataType::S16,
-                                                         DataType::U32, DataType::S32,
-                                                         DataType::F16, DataType::F32);
-
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() < 1 || input->num_dimensions() > 4,
                                     "Permutation upto 4-D input tensor is supported");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(perm.num_dimensions() < 1 || perm.num_dimensions() > 4,
diff --git a/src/core/CL/kernels/CLReverseKernel.cpp b/src/core/CL/kernels/CLReverseKernel.cpp
index 84bf5bf..796f0d0 100644
--- a/src/core/CL/kernels/CLReverseKernel.cpp
+++ b/src/core/CL/kernels/CLReverseKernel.cpp
@@ -81,20 +81,7 @@
     // Set kernel build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DNUM_REVERSE_DIMS=" + support::cpp11::to_string(axis->info()->dimension(0)));
-    switch(input->info()->element_size())
-    {
-        case 1:
-            build_opts.add_option("-DDATA_TYPE=uchar");
-            break;
-        case 2:
-            build_opts.add_option("-DDATA_TYPE=ushort");
-            break;
-        case 4:
-            build_opts.add_option("-DDATA_TYPE=uint");
-            break;
-        default:
-            ARM_COMPUTE_ERROR("Data type not supported");
-    }
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
 
     // Create kernel
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("reverse", build_opts.options()));