COMPMID-697 - Rework GEMMLowp interface on OpenCL

Reworked the interface of GemmLowp in order to make easy the integration
in Android NN

- Added support for different output stage
- Added validation for both matrix multiplication and output stage
- Added bounded relu support in the output stage
- Added in32_t bias support
- Added optimized path for vector by matrix case

This rework is required for:
- Convolution quantized
- Fully connected quantized

Change-Id: I512283d406099cf8c614dd89d0a97ed411143afc
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110625
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>

4 files changed, 523 insertions, 28 deletions

diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
index ef572cfc7e..b3227c0db9 100644
--- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
@@ -51,45 +51,88 @@ CLGEMMLowpMatrixMultiplyKernel::CLGEMMLowpMatrixMultiplyKernel()
 {
 }
 
-void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output,
-                                               int32_t a_offset, int32_t b_offset, int32_t output_offset, int32_t output_mult_int, int32_t shift)
+void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, bool is_interleaved_transposed)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::U8);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QASYMM8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
+
+    if(!is_interleaved_transposed)
+    {
+        ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1));
+    }
+
+    TensorShape in1_shape = input1->info()->tensor_shape();
+    in1_shape.collapse(2);
 
     _input0 = input0;
     _input1 = input1;
     _output = output;
 
-    // Create kernel and set static arguments
-    std::set<std::string> build_opts = { ("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0))) };
-    _kernel                          = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mm_interleaved_transposed_u8", build_opts));
-    unsigned int idx                 = 3 * num_arguments_per_2D_tensor(); //Skip the input and output parameters
-    _kernel.setArg<int32_t>(idx++, a_offset);
-    _kernel.setArg<int32_t>(idx++, b_offset);
-    _kernel.setArg<int32_t>(idx++, output_offset);
-    _kernel.setArg<int32_t>(idx++, output_mult_int);
-    _kernel.setArg<int32_t>(idx++, shift);
+    CLBuildOptions build_opts;
 
-    // Configure window
-    constexpr unsigned int num_elems_processed_per_iteration_x = 16;
-    constexpr unsigned int num_elems_processed_per_iteration_y = 4;
-    constexpr unsigned int num_elems_read_per_iteration_input0 = 4;
-    constexpr unsigned int num_elems_read_per_iteration_input1 = 16;
+    if(is_interleaved_transposed)
+    {
+        // Create kernel and set static arguments
+        build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0)));
+        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_mm_interleaved_transposed", build_opts.options()));
 
-    Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+        // Configure window
+        constexpr unsigned int num_elems_processed_per_iteration_x = 16;
+        constexpr unsigned int num_elems_processed_per_iteration_y = 4;
+        constexpr unsigned int num_elems_read_per_iteration_input0 = 4;
+        constexpr unsigned int num_elems_read_per_iteration_input1 = 16;
 
-    AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_read_per_iteration_input0, 1);
-    AccessWindowRectangle input1_access(input1->info(), 0, 0, num_elems_read_per_iteration_input1, 1);
-    AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+        Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
 
-    update_window_and_padding(win, input0_access, input1_access, output_access);
+        AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_read_per_iteration_input0, 1);
+        AccessWindowRectangle input1_access(input1->info(), 0, 0, num_elems_read_per_iteration_input1, 1);
+        AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
 
-    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
+        update_window_and_padding(win, input0_access, input1_access, output_access);
+
+        output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
+
+        ICLKernel::configure(win);
+    }
+    else
+    {
+        // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor. num_elems_processed_per_iteration_x
+        constexpr unsigned int num_elems_processed_per_iteration_x = 16;
+        const unsigned int     num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4);
+
+        build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0)));
+        build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elems_processed_per_iteration_x));
+        build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elems_processed_per_iteration_y));
+
+        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_mm", build_opts.options()));
+
+        // Configure window
+        Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+        AccessWindowStatic    input0_access(input0->info(), 0, 0, input0->info()->dimension(0), ceil_to_multiple(input0->info()->dimension(1), num_elems_processed_per_iteration_y));
+        AccessWindowStatic    input1_access(input1->info(), 0, 0, ceil_to_multiple(input1->info()->dimension(0), num_elems_processed_per_iteration_x), input1->info()->dimension(1));
+        AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+
+        update_window_and_padding(win, input0_access, input1_access, output_access);
+
+        Coordinates coord;
+        coord.set_num_dimensions(output->info()->num_dimensions());
+        output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape()));
+
+        ICLKernel::configure(win);
+    }
 
-    ICLKernel::configure(win);
+    // Set config_id for enabling LWS tuning
+    _config_id = "gemmlowp_";
+    _config_id += (is_interleaved_transposed ? "reshaped_" : "");
+    _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    _config_id += "_";
+    _config_id += (is_interleaved_transposed ? support::cpp11::to_string(input1->info()->dimension(0)) : support::cpp11::to_string(input1->info()->dimension(1)));
 }
 
 void CLGEMMLowpMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &queue)
@@ -117,7 +160,7 @@ void CLGEMMLowpMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue
         add_2D_tensor_argument(idx, _input0, slice);
         add_2D_tensor_argument(idx, _input1, slice_b);
         add_2D_tensor_argument(idx, _output, slice);
-        enqueue(queue, *this, slice);
+        enqueue(queue, *this, slice, _lws_hint);
     }
     while(window.slide_window_slice_2D(slice));
 }
diff --git a/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp b/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp
new file mode 100644
index 0000000000..96919fe3cb
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp
@@ -0,0 +1,162 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "support/ToolchainSupport.h"
+
+#include <cstddef>
+#include <cstdint>
+
+using namespace arm_compute;
+
+namespace arm_compute
+{
+class Coordinates;
+} // namespace arm_compute
+
+CLGEMMLowpOffsetContributionKernel::CLGEMMLowpOffsetContributionKernel()
+    : _vector_sum_col(nullptr), _vector_sum_row(nullptr), _mm_result(nullptr)
+{
+}
+
+void CLGEMMLowpOffsetContributionKernel::configure(ICLTensor *mm_result, const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, int32_t k, int32_t a_offset, int32_t b_offset)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32);
+
+    // Set the arguments to pass at compile time
+    CLBuildOptions build_opts;
+
+    // If a_offset == 0, vector_sum_col can be a nullptr
+    if(a_offset != 0)
+    {
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32);
+        ARM_COMPUTE_ERROR_ON(vector_sum_col->info()->dimension(0) != mm_result->info()->dimension(0));
+
+        TensorShape vector_sum_col_shape = vector_sum_col->info()->tensor_shape();
+        vector_sum_col_shape.collapse(1);
+
+        build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset));
+    }
+
+    // If b_offset == 0, vector_sum_row can be a nullptr
+    if(b_offset != 0)
+    {
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
+        ARM_COMPUTE_ERROR_ON(vector_sum_row->info()->dimension(0) != mm_result->info()->dimension(1));
+
+        TensorShape output_shape         = mm_result->info()->tensor_shape();
+        TensorShape vector_sum_row_shape = vector_sum_row->info()->tensor_shape();
+        vector_sum_row_shape.collapse(1);
+        output_shape.collapse(2);
+
+        ARM_COMPUTE_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[2], "mm_result tensor must have the same number of batches of output tensor");
+
+        if(a_offset != 0)
+        {
+            TensorShape vector_sum_col_shape = vector_sum_col->info()->tensor_shape();
+            vector_sum_col_shape.collapse(1);
+
+            ARM_COMPUTE_ERROR_ON_MSG(vector_sum_col_shape[1] != 1
+                                     && vector_sum_col_shape[1] != vector_sum_row_shape[1],
+                                     "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1");
+        }
+
+        build_opts.add_option("-DB_OFFSET=" + support::cpp11::to_string(b_offset));
+    }
+
+    build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(a_offset * b_offset * k));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_offset_contribution", build_opts.options()));
+
+    _vector_sum_col = vector_sum_col;
+    _vector_sum_row = vector_sum_row;
+    _mm_result      = mm_result;
+
+    constexpr unsigned int num_elems_processed_per_iteration = 16;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*mm_result->info(), Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal mm_result_access(mm_result->info(), 0, num_elems_processed_per_iteration);
+
+    update_window_and_padding(win, mm_result_access);
+
+    if(a_offset != 0)
+    {
+        AccessWindowHorizontal vector_sum_col_access(vector_sum_col->info(), 0, num_elems_processed_per_iteration);
+        update_window_and_padding(win, vector_sum_col_access);
+    }
+
+    if(b_offset != 0)
+    {
+        AccessWindowStatic vector_sum_row_access(vector_sum_row->info(), 0, 0, vector_sum_row->info()->dimension(0), 0);
+        update_window_and_padding(win, vector_sum_row_access);
+    }
+
+    ICLKernel::configure(win);
+}
+
+void CLGEMMLowpOffsetContributionKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+    Window slice     = collapsed.first_slice_window_3D();
+
+    // Set window for vector_sum_col
+    Window win_vector_sum_col = slice;
+    win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    // Set window for vector_sum_row
+    Window win_vector_sum_row = slice;
+    win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
+    win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _mm_result, slice);
+        if(_vector_sum_col != nullptr)
+        {
+            add_2D_tensor_argument(idx, _vector_sum_col, win_vector_sum_col);
+        }
+        if(_vector_sum_row != nullptr)
+        {
+            add_2D_tensor_argument(idx, _vector_sum_row, win_vector_sum_row);
+        }
+        enqueue(queue, *this, slice);
+    }
+    while(collapsed.slide_window_slice_3D(slice));
+}
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
new file mode 100644
index 0000000000..fa6a48e77c
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
@@ -0,0 +1,128 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+namespace arm_compute
+{
+class Coordinates;
+} // namespace arm_compute
+
+CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel()
+    : _input(nullptr), _bias(nullptr), _output(nullptr)
+{
+}
+
+void CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, int result_offset, int result_mult_int, int result_shift, int min,
+                                                              int max)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+    ARM_COMPUTE_ERROR_ON(max > 255);
+    ARM_COMPUTE_ERROR_ON(min < 0 || min > max);
+
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_ERROR_ON(bias->info()->num_dimensions() > 1);
+        ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != bias->info()->dimension(0));
+    }
+
+    _input  = input;
+    _bias   = bias;
+    _output = output;
+
+    // Set the arguments to pass at compile time
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(result_offset));
+    build_opts.add_option("-DRESULT_MULT_INT=" + support::cpp11::to_string(result_mult_int));
+    build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(result_shift));
+    build_opts.add_option_if((min != 0) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+    build_opts.add_option_if((max != 255) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
+    build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_output_stage_quantize_down", build_opts.options()));
+
+    constexpr unsigned int num_elems_processed_per_iteration = 16;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_result_access(output->info(), 0, num_elems_processed_per_iteration);
+
+    update_window_and_padding(win,
+                              input_access,
+                              output_result_access);
+
+    if(bias != nullptr)
+    {
+        AccessWindowStatic bias_access(bias->info(), 0, 0, ceil_to_multiple(bias->info()->dimension(0), num_elems_processed_per_iteration), bias->info()->tensor_shape()[1]);
+
+        update_window_and_padding(win,
+                                  bias_access);
+    }
+
+    output_result_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
+
+    ICLKernel::configure(win);
+}
+
+void CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+    Window slice     = collapsed.first_slice_window_3D();
+
+    unsigned int idx1 = num_arguments_per_3D_tensor();
+    if(_bias != nullptr)
+    {
+        Window biases_slice(slice);
+        biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
+        biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
+        add_1D_tensor_argument(idx1, _bias, biases_slice);
+    }
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx1, _output, slice);
+        enqueue(queue, *this, slice);
+    }
+    while(collapsed.slide_window_slice_3D(slice));
+}
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp b/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp
new file mode 100644
index 0000000000..6f410d3b14
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp
@@ -0,0 +1,162 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "support/ToolchainSupport.h"
+
+#include <cstddef>
+#include <cstdint>
+
+using namespace arm_compute;
+
+namespace arm_compute
+{
+class Coordinates;
+} // namespace arm_compute
+
+ICLGEMMLowpReductionKernel::ICLGEMMLowpReductionKernel()
+    : _input(), _output()
+{
+}
+
+void CLGEMMLowpMatrixAReductionKernel::configure(const ICLTensor *mtx_a, ICLTensor *vector_sum_row)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_a, 1, DataType::QASYMM8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
+
+    _input  = mtx_a;
+    _output = vector_sum_row;
+
+    // Set the arguments to pass at compile time
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(mtx_a->info()->dimension(0)));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_matrix_a_reduction", build_opts.options()));
+
+    const unsigned int num_elems_processed_per_iteration = 1;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*_output->info(), Steps(num_elems_processed_per_iteration));
+
+    AccessWindowStatic     input_access(_input->info(), 0, 0, ceil_to_multiple(_input->info()->dimension(0), 16), _input->info()->dimension(1));
+    AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration);
+
+    update_window_and_padding(win,
+                              input_access,
+                              output_access);
+
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), _output->info()->tensor_shape()));
+
+    ICLKernel::configure(win);
+}
+
+void CLGEMMLowpMatrixAReductionKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimY);
+    Window slice_in  = collapsed.first_slice_window_2D();
+    Window slice_out = collapsed.first_slice_window_2D();
+
+    // Setup input slice. Its dimensions are increased in the cl kernel.
+    slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
+    slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
+    slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice_in);
+        add_2D_tensor_argument(idx, _output, slice_out);
+        enqueue(queue, *this, slice_out);
+    }
+    while(collapsed.slide_window_slice_2D(slice_out));
+}
+
+void CLGEMMLowpMatrixBReductionKernel::configure(const ICLTensor *mtx_b, ICLTensor *vector_sum_col)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_b, 1, DataType::QASYMM8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32);
+
+    _input  = mtx_b;
+    _output = vector_sum_col;
+
+    // Set the arguments to pass at compile time
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(mtx_b->info()->dimension(0)));
+    build_opts.add_option("-DROWS_B=" + support::cpp11::to_string(mtx_b->info()->dimension(1)));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_matrix_b_reduction", build_opts.options()));
+
+    constexpr unsigned int num_elems_processed_per_iteration = 16;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*vector_sum_col->info(), Steps(num_elems_processed_per_iteration));
+
+    AccessWindowStatic     input_access(_input->info(), 0, 0, ceil_to_multiple(_input->info()->dimension(0), 16), _input->info()->dimension(1));
+    AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration);
+
+    update_window_and_padding(win,
+                              input_access,
+                              output_access);
+
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), _output->info()->tensor_shape()));
+
+    ICLKernel::configure(win);
+}
+
+void CLGEMMLowpMatrixBReductionKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window collapsed = window.collapse_if_possible(IKernel::window(), Window::DimY);
+
+    Window slice_out = collapsed.first_slice_window_2D();
+    Window slice_in  = slice_out;
+
+    slice_in.set(Window::DimY, Window::Dimension(0, 1, 1));
+    slice_in.set(Window::DimZ, Window::Dimension(0, 1, 1));
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice_in);
+        add_2D_tensor_argument(idx, _output, slice_out);
+        enqueue(queue, *this, slice_out);
+    }
+    while(collapsed.slide_window_slice_2D(slice_out));
+}