Rework OpenCL Depthwise Convolution

- Remove dedicated kernels for NCHW. Now we only use NHWC with permute
- Remove specialized kernels for 3x3 NHWC
- Simplify CLDepthwiseConvolutionLayer.cpp to call just the native
  implementation for both floating-point and quantized data types
- Develop two parametric opencl kernels for depthwise convolution layer NHWC
  (floating-point and quantized)
- Add support to export the weights to cl_image
- Extend test for depthwise convolution on opencl

Resolves COMPMID-4417

Change-Id: I253dd5d959a70783c82e62b1771a5e9f91621cb0
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5806
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>

1 files changed, 128 insertions, 95 deletions

diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
index fcfa7f878d..67ca341b0d 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
@@ -31,8 +31,10 @@
 #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/CL/CLUtils.h"
 #include "src/core/CL/CLValidate.h"
 #include "src/core/CL/ICLKernel.h"
+#include "src/core/gpu/cl/kernels/gemm/ClGemmHelpers.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
 #include "support/StringSupport.h"
@@ -41,25 +43,28 @@ namespace arm_compute
 {
 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,
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCComputeKernelInfo &dwc_info,
+                          const ConvolutionInfo &conv_info,
                           const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
-    ARM_COMPUTE_UNUSED(dwc_info);
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     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::QASYMM8_SIGNED, DataType::F16, DataType::F32);
-    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);
-    ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.depth_multiplier > 1 && dwc_info.n0 != 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().first > 1 && dwc_info.m0 != 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.dilation.x() > 1 && dwc_info.m0 != 1);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((dwc_info.export_weights_to_cl_image == true) && (export_weights_to_cl_image(weights) == false), "Export to cl_image not supported!");
+    ARM_COMPUTE_RETURN_ERROR_ON((dwc_info.export_weights_to_cl_image == true) && (conv_info.depth_multiplier > 1));
+    ARM_COMPUTE_RETURN_ERROR_ON((dwc_info.export_weights_to_cl_image == true) && ((dwc_info.n0 % 4) != 0));
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().first < 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().second < 1);
+    ARM_COMPUTE_RETURN_ERROR_ON((conv_info.dilation.x() < 1) || (conv_info.dilation.y() < 1));
     const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
     ARM_COMPUTE_UNUSED(idx_c);
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_c) != (input->dimension(idx_c) * depth_multiplier));
+    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_c) != (input->dimension(idx_c) * conv_info.depth_multiplier));
 
-    const ConvolutionInfo info{ conv_info, depth_multiplier, ActivationLayerInfo(), dilation };
-    const TensorShape     output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, info);
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info);
 
     const bool is_quantized = is_data_type_quantized(input->data_type());
 
@@ -134,111 +139,131 @@ CLDepthwiseConvolutionLayerNativeKernel::CLDepthwiseConvolutionLayerNativeKernel
       _depth_multiplier(1),
       _output_multipliers(nullptr),
       _output_shifts(nullptr),
+      _export_to_cl_image(false),
       _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,
+void CLDepthwiseConvolutionLayerNativeKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
+                                                        const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info,
                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
 {
-    configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation, output_multipliers, output_shifts);
+    configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, dwc_info, conv_info, output_multipliers, output_shifts);
 }
 
 void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext &compile_context, 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 DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info,
                                                         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,
-                                                  (output_multipliers != nullptr) ? output_multipliers->info() : nullptr, (output_shifts != nullptr) ? output_shifts->info() : nullptr));
+                                                  dwc_info, conv_info, (output_multipliers != nullptr) ? output_multipliers->info() : nullptr, (output_shifts != nullptr) ? output_shifts->info() : nullptr));
 
     auto padding_info = get_padding_info({ input, output });
 
-    const ConvolutionInfo info{ conv_info, depth_multiplier, ActivationLayerInfo(), dilation };
-    const TensorShape     output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*(input->info()), *(weights->info()), info);
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*(input->info()), *(weights->info()), conv_info);
     auto_init_if_empty(*(output->info()), input->info()->clone()->set_tensor_shape(output_shape).set_quantization_info(output->info()->quantization_info()));
 
     _input              = input;
     _output             = output;
     _weights            = weights;
     _biases             = biases;
-    _depth_multiplier   = depth_multiplier;
+    _depth_multiplier   = conv_info.depth_multiplier;
     _output_multipliers = output_multipliers;
     _output_shifts      = output_shifts;
+    _export_to_cl_image = dwc_info.export_weights_to_cl_image;
     _is_quantized       = is_data_type_quantized(input->info()->data_type());
 
-    const unsigned int n0 = adjust_vec_size(dwc_weights_info.n0, input->info()->dimension(0));
+    const unsigned int n0          = adjust_vec_size(dwc_info.n0, input->info()->dimension(0));
+    const unsigned int m0          = std::min(dwc_info.m0, (unsigned int)output->info()->dimension(1));
+    std::string        kernel_name = "";
 
     CLBuildOptions build_opts;
-    build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
-    build_opts.add_option_if(_input->info()->tensor_shape().total_size_upper(3) > 1, "-DDST_DEPTH=" + support::cpp11::to_string(static_cast<int>(_output->info()->dimension(2))));
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
-    build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(dwc_info.activation_info.activation())));
-    build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
+
+    // Update the padding for the weights tensor if we can export to cl_image
+    if(_export_to_cl_image)
+    {
+        arm_compute::opencl::kernels::gemm::update_padding_for_cl_image(weights->info());
+    }
+
+    build_opts.add_option("-cl-fast-relaxed-math");
+    build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(conv_info.act_info.activation())));
+    build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(conv_info.depth_multiplier));
+    build_opts.add_option("-DSRC_TENSOR_TYPE=BUFFER");
+    build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(_input->info()->dimension(1)));
+    build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(_input->info()->dimension(2)));
+    // Note: SRC_DATA_TYPE must have the same data type of WEI_DATA_TYPE. In quantized, we could
+    // have a case where the data types for the activation and weights are different. However, since the implementation
+    // only works when both have same data type, we have to change the offset to take into account this aspect
+    build_opts.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(_weights->info()->data_type()));
+    build_opts.add_option("-DDST_TENSOR_TYPE=BUFFER");
+    build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(1)));
+    build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(_output->info()->dimension(2)));
+    build_opts.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(_output->info()->data_type()));
+    build_opts.add_option_if_else(_export_to_cl_image, "-DWEI_TENSOR_TYPE=IMAGE", "-DWEI_TENSOR_TYPE=BUFFER");
+    build_opts.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights->info()->dimension(1)));
+    build_opts.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(weights->info()->dimension(2)));
+    build_opts.add_option("-DWEI_DATA_TYPE=" + get_cl_type_from_data_type(weights->info()->data_type()));
+    build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_stride_info.pad_top()));
+    build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_stride_info.pad_left()));
+    build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.pad_stride_info.stride().first));
+    build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.pad_stride_info.stride().second));
+    build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(conv_info.dilation.x()));
+    build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(conv_info.dilation.y()));
     build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
-    build_opts.add_option("-DSRC_DIM1=" + support::cpp11::to_string(_input->info()->dimension(1)));
-    build_opts.add_option("-DSRC_DIM2=" + support::cpp11::to_string(_input->info()->dimension(2)));
-    build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(weights->info()->dimension(1)));
-    build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(weights->info()->dimension(2)));
-    build_opts.add_option("-DCONV_PAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
-    build_opts.add_option("-DCONV_PAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
-    build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
-    build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
-    build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
-    build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
-    build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(_input->info()->dimension(0) % n0));
-
-    std::string kernel_name = (_is_quantized) ? "dwc_MxN_native_quantized8_nhwc" : "dwc_MxN_native_fp_nhwc";
+    build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
+    build_opts.add_option("-DM0_A=" + support::cpp11::to_string(weights->info()->dimension(1) + m0 - 1));
+    build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(_input->info()->dimension(0) % n0));
+    build_opts.add_option_if(_input->info()->num_dimensions() > 3, "-DBATCHED_EXECUTION");
+    if(biases != nullptr)
+    {
+        build_opts.add_option(std::string("-DHAS_BIAS"));
+        build_opts.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(biases->info()->data_type())));
+    }
 
     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();
+        kernel_name                          = "dwc_native_quantized_nhwc";
+        const UniformQuantizationInfo iqinfo = input->info()->quantization_info().uniform();
+        const UniformQuantizationInfo wqinfo = weights->info()->quantization_info().uniform();
+        const UniformQuantizationInfo oqinfo = output->info()->quantization_info().uniform();
 
-        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_if(is_data_type_quantized_per_channel(weights->info()->data_type()), "-DPER_CHANNEL_QUANTIZATION");
+        PixelValue zero_value = PixelValue(0, input->info()->data_type(), input->info()->quantization_info());
+        int        zero_value_s32;
+        zero_value.get(zero_value_s32);
 
-        // Compute non-per-channel multiplier and shift anyway to make OpenCL kernel simpler
-        float multiplier        = iq_info.scale * wq_info.scale / oq_info.scale;
+        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_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
-        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
-
-        if(dwc_info.activation_info.enabled())
-        {
-            int a_val{};
-            int b_val{};
-            std::tie(b_val, a_val) = get_quantized_activation_min_max(dwc_info.activation_info, input->info()->data_type(), oq_info);
-
-            const int o1 = oq_info.offset;
-
-            build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
-            build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
-            build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
-
-            const float s1 = iq_info.scale;
-            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("-DDST_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
+        build_opts.add_option("-DDST_SHIFT=" + support::cpp11::to_string(output_shift));
+        build_opts.add_option("-DSRC_OFFSET=" + support::cpp11::to_string(-iqinfo.offset));
+        build_opts.add_option("-DWEI_OFFSET=" + support::cpp11::to_string(-wqinfo.offset));
+        build_opts.add_option("-DDST_OFFSET=" + support::cpp11::to_string(oqinfo.offset));
+        build_opts.add_option("-DZERO_VALUE=" + support::cpp11::to_string(zero_value_s32));
+        build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(DataType::S32));
+        build_opts.add_option("-DDST_MULTIPLIERS_DATA_TYPE=" + get_cl_type_from_data_type(_output_multipliers->info()->data_type()));
+        build_opts.add_option("-DDST_SHIFTS_DATA_TYPE=" + get_cl_type_from_data_type(_output_shifts->info()->data_type()));
+        build_opts.add_option_if_else(weights->info()->data_type() == DataType::QSYMM8_PER_CHANNEL, "-DQUANTIZATION_TYPE=PER_CHANNEL", "-DQUANTIZATION_TYPE=PER_TENSOR");
+        // Note: We expect the input and output tensors to always adopt a per-tensor quantization approach
+        int a_val{};
+        int b_val{};
+        std::tie(b_val, a_val) = get_quantized_activation_min_max(conv_info.act_info, input->info()->data_type(), oqinfo);
+
+        build_opts.add_option_if(conv_info.act_info.enabled(), "-DA_VAL=" + support::cpp11::to_string(a_val));
+        build_opts.add_option_if(conv_info.act_info.enabled(), "-DB_VAL=" + support::cpp11::to_string(b_val));
     }
     else
     {
-        build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.a()));
-        build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.b()));
+        kernel_name = "dwc_native_fp_nhwc";
+        build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+        build_opts.add_option("-DZERO_VALUE=" + support::cpp11::to_string(0));
+        build_opts.add_option_if(conv_info.act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(conv_info.act_info.a()));
+        build_opts.add_option_if(conv_info.act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(conv_info.act_info.b()));
     }
 
-    Window win = calculate_max_window(*(output->info()), Steps(n0));
+    Window win = calculate_max_window(*(output->info()), Steps(n0, m0));
     ICLKernel::configure_internal(win);
 
     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
@@ -264,10 +289,9 @@ void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext &
 }
 
 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 ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
+                                                         const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info, 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, output_multipliers, output_shifts));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, dwc_info, conv_info, output_multipliers, output_shifts));
     return Status{};
 }
 
@@ -278,37 +302,46 @@ void CLDepthwiseConvolutionLayerNativeKernel::run(const Window &window, cl::Comm
 
     // Collapse window
     Window window_collapsed = window.collapse(ICLKernel::window(), Window::DimZ);
-    Window slice_in         = window.first_slice_window_4D();
-    Window slice_out        = window_collapsed.first_slice_window_4D();
+
+    Window slice = window_collapsed.first_slice_window_4D();
 
     if(_depth_multiplier != 1)
     {
-        ARM_COMPUTE_ERROR_ON(slice_out.x().step() != 1);
-        slice_out.set(Window::DimX, Window::Dimension(0, _input->info()->tensor_shape()[0], 1));
+        // If the depth multiplier > 1, we need to use the input channels rather than the output channels
+        ARM_COMPUTE_ERROR_ON(slice.x().step() != 1);
+        slice.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();
+    cl::Image2D weights_cl_image;
 
-    // Set output multipliers in case of quantized data type
-    if(_is_quantized)
+    if(_export_to_cl_image)
     {
-        add_1D_tensor_argument(idx, _output_multipliers, slice_in);
-        add_1D_tensor_argument(idx, _output_shifts, slice_in);
+        const size_t      image_w = _weights->info()->dimension(0) / 4;
+        const size_t      image_h = _weights->info()->dimension(1) * _weights->info()->dimension(2) * _weights->info()->dimension(3);
+        const TensorShape shape2d(image_w, image_h);
+        const size_t      image_row_pitch = _weights->info()->strides_in_bytes()[1];
+
+        // Export cl_buffer to cl_image
+        weights_cl_image = create_image2d_from_buffer(CLKernelLibrary::get().context(), _weights->cl_buffer(), shape2d, _weights->info()->data_type(), image_row_pitch);
     }
 
-    if(_biases != nullptr)
+    unsigned int idx = 0;
+    add_4D_tensor_argument(idx, _input, slice);
+    add_4D_tensor_argument(idx, _output, slice);
+    if(_export_to_cl_image)
     {
-        add_1D_tensor_argument(idx, _biases, slice_in);
+        _kernel.setArg(idx++, weights_cl_image);
     }
-
-    do
+    add_4D_tensor_argument(idx, _weights, slice);
+    if(_is_quantized)
+    {
+        add_1D_tensor_argument(idx, _output_multipliers, slice);
+        add_1D_tensor_argument(idx, _output_shifts, slice);
+    }
+    if(_biases != nullptr)
     {
-        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);
-        enqueue(queue, *this, slice_out, lws_hint());
+        add_1D_tensor_argument(idx, _biases, slice);
     }
-    while(window_collapsed.slide_window_slice_4D(slice_out) && window.slide_window_slice_4D(slice_in));
+    enqueue(queue, *this, slice, lws_hint());
 }
 } // namespace arm_compute