Make ClDirectConvolutionKernel stateless

ClDirectorConvolution triggers ClActivation (if enabled)

Remove static tuner as the interface need to be changed base on new api. Remove functions in ClScaleKernel specific for static Tuner.

Solves: COMPMID-4010

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I7861c3462fda323a6fe1891834068a462245cb1b
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5262
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 694 insertions, 36 deletions

diff --git a/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.cpp b/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.cpp
new file mode 100644
index 0000000000..f071dbc468
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.cpp
@@ -0,0 +1,587 @@
+/*
+ * Copyright (c) 2017-2021 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 "src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/PixelValue.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
+#include "src/core/AccessWindowStatic.h"
+#include "src/core/CL/CLValidate.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+#include "support/Cast.h"
+#include "support/StringSupport.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8_SIGNED, DataType::QASYMM8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights);
+
+    const DataLayout data_layout = src->data_layout();
+    const int        width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int        height_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+    const int        channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != weights->dimension(height_idx), "Weights should have same width and height");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(channel_idx) != src->dimension(channel_idx),
+                                    "Weights feature map dimension should match the respective src's one");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4, "Weights can be at most 4 dimensional");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 1) && std::get<0>(conv_info.stride()) > 3, "Strides larger than 3 not supported for 1x1 convolution.");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 3 || weights->dimension(width_idx) == 5 || weights->dimension(width_idx) == 9)
+                                    && std::get<0>(conv_info.stride()) > 2,
+                                    "Strides larger than 2 not supported for 3x3, 5x5, 9x9 convolution.");
+
+    if(data_layout == DataLayout::NCHW)
+    {
+        if(is_data_type_quantized(src->data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != 1 && weights->dimension(width_idx) != 3 && weights->dimension(width_idx) != 5 && weights->dimension(width_idx) != 9,
+                                            "Kernel sizes other than 1x1, 3x3, 5x5 or 9x9 are not supported with quantized data types");
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != 1 && weights->dimension(width_idx) != 3 && weights->dimension(width_idx) != 5,
+                                            "Kernel sizes other than 1x1, 3x3 or 5x5 are not supported with float data types");
+        }
+    }
+
+    if(biases != nullptr)
+    {
+        if(is_data_type_quantized_asymmetric(src->data_type()))
+        {
+            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_MSG(biases->dimension(0) != weights->dimension(3),
+                                        "Biases size and number of src feature maps should match");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->num_dimensions() > 1,
+                                        "Biases should be one dimensional");
+    }
+
+    // Checks performed when dst is configured
+    if(dst->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(),
+                                                           misc::shape_calculator::compute_deep_convolution_shape(*src, *weights, conv_info));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+    }
+
+    const auto data_type = src->data_type();
+    if(is_data_type_quantized(data_type))
+    {
+        const UniformQuantizationInfo iqinfo = src->quantization_info().uniform();
+        const UniformQuantizationInfo wqinfo = weights->quantization_info().uniform();
+        const UniformQuantizationInfo oqinfo = dst->quantization_info().uniform();
+
+        float multiplier        = iqinfo.scale * wqinfo.scale / oqinfo.scale;
+        int   output_multiplier = 0;
+        int   output_shift      = 0;
+        ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
+    }
+    return Status{};
+}
+
+inline bool can_run_optimized_kernel_for_bifrost_nchw(GPUTarget gpu_target, unsigned int conv_stride_x, unsigned int conv_stride_y, unsigned int kernel_size,
+                                                      DataType data_type, DataLayout data_layout)
+{
+    return gpu_target_is_in(gpu_target,
+                            GPUTarget::G71, GPUTarget::G72, GPUTarget::G76,
+                            GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT,
+                            GPUTarget::G52, GPUTarget::G52LIT)
+           && (kernel_size <= 5)
+           && (conv_stride_x == 1) && (conv_stride_y == 1)
+           && (data_type == DataType::F32)
+           && (data_layout == DataLayout::NCHW);
+}
+
+inline void setup_num_elems_nchw(unsigned int &num_elems_read_per_iteration_x, unsigned int &num_elems_read_per_iteration_y,
+                                 unsigned int &num_elems_written_per_iteration_x, unsigned int &num_elems_written_per_iteration_y,
+                                 unsigned int kernel_size, const PadStrideInfo &conv_info, const GPUTarget target, ITensorInfo *src)
+{
+    const DataType   data_type     = src->data_type();
+    const DataLayout data_layout   = src->data_layout();
+    unsigned int     conv_stride_x = std::get<0>(conv_info.stride());
+    unsigned int     conv_stride_y = std::get<1>(conv_info.stride());
+
+    const bool run_optimized_bifrost = can_run_optimized_kernel_for_bifrost_nchw(target, conv_stride_x, conv_stride_y, kernel_size, data_type, data_layout);
+
+    if(run_optimized_bifrost)
+    {
+        // Configure kernel window
+        switch(kernel_size)
+        {
+            case 1:
+            {
+                num_elems_read_per_iteration_x    = 4;
+                num_elems_read_per_iteration_y    = 4;
+                num_elems_written_per_iteration_x = 4;
+                num_elems_written_per_iteration_y = 4;
+                break;
+            }
+            case 3:
+            {
+                num_elems_read_per_iteration_x    = 6;
+                num_elems_read_per_iteration_y    = 5;
+                num_elems_written_per_iteration_x = 4;
+                num_elems_written_per_iteration_y = 3;
+                break;
+            }
+            case 5:
+            {
+                num_elems_read_per_iteration_x    = 8;
+                num_elems_read_per_iteration_y    = 6;
+                num_elems_written_per_iteration_x = 4;
+                num_elems_written_per_iteration_y = 2;
+                break;
+            }
+            default:
+            {
+                ARM_COMPUTE_ERROR("Kernel size not optimized for Bifrost");
+            }
+        }
+    }
+    else
+    {
+        num_elems_read_per_iteration_y    = kernel_size;
+        num_elems_written_per_iteration_x = 8;
+        num_elems_written_per_iteration_y = 1;
+        switch(kernel_size)
+        {
+            case 1:
+                switch(conv_stride_x)
+                {
+                    case 1:
+                        num_elems_read_per_iteration_x = 8;
+                        break;
+                    case 2:
+                        num_elems_read_per_iteration_x = 16;
+                        break;
+                    case 3:
+                        switch(src->element_size())
+                        {
+                            case 1:
+                                num_elems_read_per_iteration_x = 28;
+                                break;
+                            case 2:
+                                num_elems_read_per_iteration_x = 24;
+                                break;
+                            case 4:
+                                num_elems_read_per_iteration_x = 22;
+                                break;
+                            default:
+                                ARM_COMPUTE_ERROR("Invalid data size");
+                        }
+                        break;
+                    default:
+                        ARM_COMPUTE_ERROR("Invalid convolution stride X");
+                }
+                break;
+            case 3:
+                switch(conv_stride_x)
+                {
+                    case 1:
+                        num_elems_read_per_iteration_x = 10;
+                        break;
+                    case 2:
+                        num_elems_read_per_iteration_x = 17;
+                        break;
+                    default:
+                        ARM_COMPUTE_ERROR("Invalid convolution stride X");
+                }
+                break;
+            case 5:
+                switch(conv_stride_x)
+                {
+                    case 1:
+                        num_elems_read_per_iteration_x = 12;
+                        break;
+                    case 2:
+                        num_elems_read_per_iteration_x = 20;
+                        break;
+                    default:
+                        ARM_COMPUTE_ERROR("Invalid convolution stride X");
+                }
+                break;
+            case 9:
+                switch(conv_stride_x)
+                {
+                    case 1:
+                        num_elems_read_per_iteration_x = 16;
+                        break;
+                    case 2:
+                        num_elems_read_per_iteration_x = 24;
+                        break;
+                    default:
+                        ARM_COMPUTE_ERROR("Invalid convolution stride X");
+                }
+                break;
+            default:
+                ARM_COMPUTE_ERROR("Invalid direct convolution size");
+        }
+    }
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITensorInfo *weights, ITensorInfo *dst, const PadStrideInfo &conv_info, const GPUTarget target)
+{
+    const DataLayout data_layout = src->data_layout();
+
+    // Get dst shape
+    TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *weights, conv_info);
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*dst, output_shape,
+                       1,
+                       src->data_type(),
+                       src->quantization_info());
+
+    if(data_layout == DataLayout::NHWC)
+    {
+        const unsigned int vec_size = std::min(static_cast<unsigned int>(dst->tensor_shape()[0]), 4u);
+
+        // Create window and update padding
+        Window win = calculate_max_window(*dst, Steps(vec_size, 1U));
+        dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape()));
+        Status err = Status{};
+        return std::make_pair(err, win);
+    }
+    else if(data_layout == DataLayout::NCHW)
+    {
+        const int          width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+        const unsigned int kernel_size = weights->dimension(width_idx);
+
+        unsigned int num_elems_read_per_iteration_x    = 0;
+        unsigned int num_elems_read_per_iteration_y    = 0;
+        unsigned int num_elems_written_per_iteration_x = 0;
+        unsigned int num_elems_written_per_iteration_y = 0;
+
+        unsigned int conv_pad_left = conv_info.pad_left();
+        unsigned int conv_pad_top  = conv_info.pad_top();
+        unsigned int conv_stride_x = std::get<0>(conv_info.stride());
+        unsigned int conv_stride_y = std::get<1>(conv_info.stride());
+
+        setup_num_elems_nchw(num_elems_read_per_iteration_x, num_elems_read_per_iteration_y,
+                             num_elems_written_per_iteration_x, num_elems_written_per_iteration_y,
+                             kernel_size, conv_info, target, src);
+
+        // Create window and update padding
+        bool   window_changed = false;
+        Window win            = calculate_max_window(*dst, Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y));
+
+        AccessWindowRectangle input_access(src, -conv_pad_left, -conv_pad_top, num_elems_read_per_iteration_x, num_elems_read_per_iteration_y, conv_stride_x, conv_stride_y);
+        AccessWindowStatic    weights_access(weights, 0, 0, kernel_size, kernel_size);
+        AccessWindowRectangle output_access(dst, 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y);
+        window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
+        output_access.set_valid_region(win, ValidRegion(Coordinates(), dst->tensor_shape()));
+        Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+        return std::make_pair(err, win);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not supported");
+    }
+}
+} // namespace
+
+BorderSize ClDirectConvolutionKernel::border_size() const
+{
+    return _border_size;
+}
+
+void ClDirectConvolutionKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst,
+                                          const PadStrideInfo &conv_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst);
+
+    // Perform validation
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src,
+                                                  weights,
+                                                  (biases != nullptr) ? biases : nullptr,
+                                                  dst,
+                                                  conv_info));
+
+    const int conv_stride_x = std::get<0>(conv_info.stride());
+    const int conv_stride_y = std::get<1>(conv_info.stride());
+
+    _data_layout = src->data_layout();
+    _conv_info   = conv_info;
+
+    const unsigned int width_idx   = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+    const unsigned int height_idx  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+    const unsigned int channel_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
+    const unsigned int kernel_size = weights->dimension(width_idx);
+    const DataType     data_type   = src->data_type();
+
+    const GPUTarget gpu_target = get_target();
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(src, weights, dst, conv_info, gpu_target);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
+    std::stringstream kernel_name;
+    CLBuildOptions    build_options;
+
+    if(_data_layout == DataLayout::NHWC)
+    {
+        _border_size = BorderSize();
+
+        kernel_name << "direct_convolution_nhwc";
+
+        const unsigned int n0               = win_config.second.x().step();
+        const unsigned int m0               = win_config.second.y().step();
+        const unsigned int k0               = adjust_vec_size(16u, src->dimension(channel_idx));
+        const unsigned int partial_store_n0 = dst->dimension(channel_idx) % n0;
+        const unsigned int partial_store_m0 = (dst->dimension(width_idx) * dst->dimension(height_idx)) % m0;
+        const unsigned int pad_left         = conv_info.pad_left();
+        const unsigned int pad_top          = conv_info.pad_top();
+
+        if(biases != nullptr)
+        {
+            build_options.add_option(std::string("-DHAS_BIAS"));
+            build_options.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(biases->data_type())));
+        }
+        build_options.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(width_idx)));
+        build_options.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(height_idx)));
+        build_options.add_option("-DSRC_CHANNELS=" + support::cpp11::to_string(src->dimension(channel_idx)));
+        build_options.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
+        build_options.add_option("-DDST_WIDTH=" + support::cpp11::to_string(dst->dimension(width_idx)));
+        build_options.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(height_idx)));
+        build_options.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst->dimension(channel_idx)));
+        build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst->data_type()));
+        build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights->dimension(width_idx)));
+        build_options.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(weights->dimension(height_idx)));
+        build_options.add_option("-DWEI_DATA_TYPE=" + get_cl_type_from_data_type(weights->data_type()));
+        build_options.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_stride_x));
+        build_options.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_stride_y));
+        build_options.add_option("-DPAD_LEFT=" + support::cpp11::to_string(pad_left));
+        build_options.add_option("-DPAD_TOP=" + support::cpp11::to_string(pad_top));
+        build_options.add_option("-DN0=" + support::cpp11::to_string(n0));
+        build_options.add_option("-DM0=" + support::cpp11::to_string(m0));
+        build_options.add_option("-DK0=" + support::cpp11::to_string(k0));
+        build_options.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
+        build_options.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
+
+        if(is_data_type_quantized(data_type))
+        {
+            const UniformQuantizationInfo iqinfo = src->quantization_info().uniform();
+            const UniformQuantizationInfo wqinfo = weights->quantization_info().uniform();
+            const UniformQuantizationInfo oqinfo = dst->quantization_info().uniform();
+
+            PixelValue zero_value = PixelValue(0, src->data_type(), src->quantization_info());
+            int        zero_value_s32;
+            zero_value.get(zero_value_s32);
+
+            float multiplier        = iqinfo.scale * wqinfo.scale / oqinfo.scale;
+            int   output_multiplier = 0;
+            int   output_shift      = 0;
+            quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
+            build_options.add_option("-DIS_QUANTIZED");
+            build_options.add_option("-DDST_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
+            build_options.add_option("-DDST_SHIFT=" + support::cpp11::to_string(output_shift));
+            build_options.add_option("-DSRC_OFFSET=" + support::cpp11::to_string(-iqinfo.offset));
+            build_options.add_option("-DWEI_OFFSET=" + support::cpp11::to_string(-wqinfo.offset));
+            build_options.add_option("-DDST_OFFSET=" + support::cpp11::to_string(oqinfo.offset));
+            build_options.add_option("-DZERO_VALUE=" + support::cpp11::to_string(zero_value_s32));
+            build_options.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(DataType::S32));
+        }
+        else
+        {
+            build_options.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(data_type));
+            build_options.add_option("-DSRC_OFFSET=" + support::cpp11::to_string(0));
+            build_options.add_option("-DWEI_OFFSET=" + support::cpp11::to_string(0));
+            build_options.add_option("-DDST_OFFSET=" + support::cpp11::to_string(0));
+        }
+    }
+    else
+    {
+        _border_size = BorderSize(src->padding());
+
+        kernel_name << "direct_convolution" << kernel_size << "x" << kernel_size;
+
+        build_options.add_option_if(biases != nullptr, std::string("-DHAS_BIAS"));
+
+        const bool run_optimized_for_bifrost = can_run_optimized_kernel_for_bifrost_nchw(gpu_target, conv_stride_x, conv_stride_y, kernel_size, data_type, _data_layout);
+
+        if(run_optimized_for_bifrost)
+        {
+            build_options.add_option(std::string("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(weights->dimension(channel_idx))));
+
+            kernel_name << "_f32_bifrost";
+        }
+        else
+        {
+            build_options.add_option(std::string("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)));
+            build_options.add_option(std::string("-DDATA_SIZE=" + get_data_size_from_data_type(data_type)));
+            build_options.add_option(std::string("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(weights->dimension(channel_idx))));
+            build_options.add_option(std::string("-DSTRIDE_X=" + support::cpp11::to_string(conv_stride_x)));
+            build_options.add_option(std::string("-DDATA_TYPE_PROMOTED=" + get_cl_type_from_data_type(data_type)));
+
+            if(is_data_type_quantized(data_type))
+            {
+                const UniformQuantizationInfo iqinfo = src->quantization_info().uniform();
+                const UniformQuantizationInfo wqinfo = weights->quantization_info().uniform();
+                const UniformQuantizationInfo oqinfo = dst->quantization_info().uniform();
+
+                float multiplier        = iqinfo.scale * wqinfo.scale / oqinfo.scale;
+                int   output_multiplier = 0;
+                int   output_shift      = 0;
+                quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
+                build_options.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
+                build_options.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
+                build_options.add_option("-DKERNEL_SIZE=" + support::cpp11::to_string(kernel_size));
+                build_options.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-iqinfo.offset));
+                build_options.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-wqinfo.offset));
+                build_options.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(oqinfo.offset));
+
+                kernel_name.str("direct_convolution_quantized");
+            }
+        }
+    }
+
+    _kernel = create_kernel(compile_context, kernel_name.str(), build_options.options());
+
+    // Set config_id for enabling LWS tuning
+    _config_id = kernel_name.str();
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_type(data_type));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(kernel_size);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(border_size().left);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(border_size().top);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(border_size().right);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(border_size().bottom);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(conv_stride_x);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(conv_stride_y);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(dst->dimension(width_idx));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(dst->dimension(height_idx));
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_layout(_data_layout));
+}
+
+Status ClDirectConvolutionKernel::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info,
+                                           const GPUTarget target)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, conv_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), weights->clone().get(), dst->clone().get(), conv_info, target).first);
+
+    return Status{};
+}
+
+void ClDirectConvolutionKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    // Get initial windows
+    Window slice = window.first_slice_window_3D();
+
+    const auto src     = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
+    const auto weights = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
+    const auto biases  = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
+    auto       dst     = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+
+    if(_data_layout == DataLayout::NHWC)
+    {
+        slice.set(Window::DimY, Window::Dimension(0, dst->info()->dimension(1) * dst->info()->dimension(2), 1));
+        slice.set(Window::DimZ, Window::Dimension(0, dst->info()->dimension(3), 1));
+
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, src, slice);
+        add_3D_tensor_argument(idx, dst, slice);
+        add_3D_tensor_argument(idx, weights, slice);
+        if(biases != nullptr)
+        {
+            add_1D_tensor_argument(idx, biases, slice);
+        }
+        _kernel.setArg(idx++, static_cast<unsigned int>(weights->info()->strides_in_bytes()[3]));
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    else
+    {
+        Window win_in = window;
+
+        win_in.adjust(Window::DimX, -_conv_info.pad_left(), true);
+        win_in.adjust(Window::DimY, -_conv_info.pad_top(), true);
+
+        const int width_idx  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+        const int height_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+
+        const int conv_stride_x = std::get<0>(_conv_info.stride());
+        const int conv_stride_y = std::get<1>(_conv_info.stride());
+
+        win_in.set_dimension_step(width_idx, window[width_idx].step() * conv_stride_x);
+        win_in.set_dimension_step(height_idx, window[height_idx].step() * conv_stride_y);
+
+        Window       slice_in = win_in.first_slice_window_3D();
+        unsigned int idx1     = 2 * num_arguments_per_3D_tensor();
+        add_3D_tensor_argument(idx1, weights, slice);
+
+        if(biases != nullptr)
+        {
+            Window slice_biases;
+            slice_biases.use_tensor_dimensions(biases->info()->tensor_shape());
+            add_1D_tensor_argument(idx1, biases, slice_biases);
+        }
+
+        _kernel.setArg(idx1++, static_cast<unsigned int>(weights->info()->strides_in_bytes()[3]));
+
+        do
+        {
+            unsigned int idx = 0;
+            add_3D_tensor_argument(idx, src, slice_in);
+            add_3D_tensor_argument(idx, dst, slice);
+            enqueue(queue, *this, slice, lws_hint());
+        }
+        while(window.slide_window_slice_3D(slice) && win_in.slide_window_slice_3D(slice_in));
+    }
+}
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h b/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h
new file mode 100644
index 0000000000..ff2f5619db
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2017-2021 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_CL_DIRECT_CONVOLUTION_KERNEL_H
+#define ARM_COMPUTE_CL_DIRECT_CONVOLUTION_KERNEL_H
+
+#include "src/core/common/Macros.h"
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+/** Interface for the  direct convolution kernel.
+ */
+class ClDirectConvolutionKernel : public ICLKernel
+{
+public:
+    ClDirectConvolutionKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClDirectConvolutionKernel);
+    /** Set the src, weights, biases and dst tensors info.
+     *
+     * @note: Due to set_valid_region(), thus src/weights/biases cannot be const. Need to change this once the set_valid_region() is removed.
+     *
+     * @note: DirectConvolution only works in the following configurations:
+     *        1x1 convolution with stride_x = 1/2/3, stride_y = 1/2/3
+     *        3x3 convolution with stride_x = 1/2, stride_y = 1/2
+     *        5x5 convolution with stride_x = 1/2, stride_y = 1/2
+     *        9x9 convolution with stride_x = 1/2, stride_y = 1/2
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src             The src tensor info to convolve. 3 lower dimensions represent a single src [width, height, IFM],
+     *                             while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8_SIGNED/QASYMM8/F16/F32.
+     * @param[in]  weights         Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
+     *                             The 3rd dimension must be the same as the src's volume 3rd dimension.
+     *                             Data type supported:Same as @p src.
+     * @param[in]  biases          Biases tensor info. Biases are 1D tensor with dimension [OFM].
+     *                             Data type supported: Should match @p src data type, except for src of QASYMM8 and QASYMM8_SIGNED type where biases should be of S32 type
+     * @param[out] dst             Output tensor info.
+     *                             The 3rd dimensions must be equal to the 4th dimension of the @p kernels tensor. Data types supported: Same as @p src.
+     * @param[in]  conv_info       Contains padding and stride information described in @ref PadStrideInfo.
+     */
+    void configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info);
+    /** Static function to check if given info will lead to a valid configuration of @ref ClDirectConvolutionKernel
+     *
+     * @param[in] src       The src tensor info to convolve. 3 lower dimensions represent a single src [width, height, IFM],
+     *                      while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8_SIGNED/QASYMM8/F16/F32.
+     * @param[in] weights   Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
+     *                      The 3rd dimension must be the same as the src's volume 3rd dimension.
+     *                      Data type supported:Same as @p src.
+     * @param[in] biases    Biases tensor info. Biases are 1D tensor with dimension [OFM].
+     *                      Data type supported: Should match @p src data type, except for src of QASYMM8 and QASYMM8_SIGNED type where biases should be of S32 type.
+     * @param[in] dst       Output tensor info.
+     *                      The 3rd dimensions must be equal to the 4th dimension of the @p kernels tensor. Data types supported: Same as @p src.
+     * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in] target    Target GPU architecture.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, const GPUTarget target);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
+
+public:
+    DataLayout    _data_layout{};
+    BorderSize    _border_size{};
+    PadStrideInfo _conv_info{};
+};
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CL_DIRECT_CONVOLUTION_KERNEL_H */
diff --git a/src/core/gpu/cl/kernels/ClScaleKernel.cpp b/src/core/gpu/cl/kernels/ClScaleKernel.cpp
index 0882f29135..7fb5d2a5d3 100644
--- a/src/core/gpu/cl/kernels/ClScaleKernel.cpp
+++ b/src/core/gpu/cl/kernels/ClScaleKernel.cpp
@@ -146,28 +146,26 @@ void ClScaleKernel::configure(const CLCompileContext &compile_context, ITensorIn
     auto padding_info = get_padding_info({ src, dst });
 
     // Info required for the static tuning
-    _info        = info;
-    _data_type   = src->data_type();
-    _data_layout = _info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : _info.data_layout;
+    _data_layout = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout;
 
     float wr = 0.f;
     float hr = 0.f;
-    std::tie(wr, hr) = calculate_scale_factors(src, dst, _data_layout, _info.align_corners);
-    const bool call_quantized_kernel = is_data_type_quantized_asymmetric(src->data_type()) && _info.interpolation_policy == InterpolationPolicy::BILINEAR;
+    std::tie(wr, hr) = calculate_scale_factors(src, dst, _data_layout, info.align_corners);
+    const bool call_quantized_kernel = is_data_type_quantized_asymmetric(src->data_type()) && info.interpolation_policy == InterpolationPolicy::BILINEAR;
 
     // Compute actual border size
     BorderSize border  = border_size();
     const bool is_nhwc = _data_layout == DataLayout::NHWC;
 
     // Area interpolation behaves as Nearest Neighbour in case of up-sampling
-    auto interpolation_policy_to_use = _info.interpolation_policy;
-    if(_info.interpolation_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f)
+    auto interpolation_policy_to_use = info.interpolation_policy;
+    if(info.interpolation_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f)
     {
         interpolation_policy_to_use = InterpolationPolicy::NEAREST_NEIGHBOR;
     }
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(src, dst, _info, border);
+    auto win_config = validate_and_configure_window(src, dst, info, border);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure_internal(win_config.second);
 
@@ -178,7 +176,7 @@ void ClScaleKernel::configure(const CLCompileContext &compile_context, ITensorIn
     build_opts.add_option("-DBORDER_SIZE=" + support::cpp11::to_string(border.right));
     build_opts.add_option_if(info.border_mode == BorderMode::REPLICATE, "-DBORDER_MODE_REPLICATE");
     build_opts.add_option_if(is_nhwc, "-DDEPTH_OUT=" + support::cpp11::to_string(dst->dimension(2)));
-    build_opts.add_option_if_else(_info.sampling_policy == SamplingPolicy::CENTER, "-DSAMPLING_POLICY_CENTER", "-DSAMPLING_POLICY_TOP_LEFT");
+    build_opts.add_option_if_else(info.sampling_policy == SamplingPolicy::CENTER, "-DSAMPLING_POLICY_CENTER", "-DSAMPLING_POLICY_TOP_LEFT");
     build_opts.add_option_if(info.align_corners, "-DALIGN_CORNERS");
     if(call_quantized_kernel)
     {
@@ -209,13 +207,10 @@ void ClScaleKernel::configure(const CLCompileContext &compile_context, ITensorIn
     _kernel.setArg<float>(idx++, wr);
     _kernel.setArg<float>(idx++, hr);
 
-    // Set to enable static tuning
-    _output_x_dim = dst->dimension(0);
-
     // Set config_id for enabling LWS tuning
     _config_id = "scale_";
-    _config_id += (_info.border_mode == BorderMode::REPLICATE ? "Bord_rep" : "");
-    _config_id += (_info.sampling_policy == SamplingPolicy::CENTER ? "center" : "topleft");
+    _config_id += (info.border_mode == BorderMode::REPLICATE ? "Bord_rep" : "");
+    _config_id += (info.sampling_policy == SamplingPolicy::CENTER ? "center" : "topleft");
     _config_id += (is_nhwc ? "nhwc" : "nchw");
     _config_id += "_";
     _config_id += support::cpp11::to_string(dst->dimension(0));
diff --git a/src/core/gpu/cl/kernels/ClScaleKernel.h b/src/core/gpu/cl/kernels/ClScaleKernel.h
index b6eea0620b..10a1105f08 100644
--- a/src/core/gpu/cl/kernels/ClScaleKernel.h
+++ b/src/core/gpu/cl/kernels/ClScaleKernel.h
@@ -70,29 +70,8 @@ public:
     BorderSize border_size() const override;
     void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
 
-    // Getter for interpolation policy
-    InterpolationPolicy get_interpolation_policy() const
-    {
-        return _info.interpolation_policy;
-    }
-
-    // Getter for data type
-    DataType get_data_type() const
-    {
-        return _data_type;
-    }
-
-    // Getter for output x dimension
-    unsigned int get_output_x_dim() const
-    {
-        return _output_x_dim;
-    }
-
 private:
-    ScaleKernelInfo _info         = ScaleKernelInfo(InterpolationPolicy::NEAREST_NEIGHBOR, BorderMode::UNDEFINED);
-    DataType        _data_type    = DataType::UNKNOWN;
-    DataLayout      _data_layout  = DataLayout::UNKNOWN;
-    unsigned int    _output_x_dim = 0;
+    DataLayout _data_layout = DataLayout::UNKNOWN;
 };
 } // namespace kernels
 } // namespace opencl