COMPMID-1727 - CL: Implement Gather

Change-Id: I3d859da09a4de1019bb8c2046725eab942247927
Reviewed-on: https://review.mlplatform.org/386
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

14 files changed, 1073 insertions, 4 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index cfcfb7400b..e68769b6ae 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -83,6 +83,7 @@
 #include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h"
+#include "arm_compute/core/CL/kernels/CLGatherKernel.h"
 #include "arm_compute/core/CL/kernels/CLGaussian3x3Kernel.h"
 #include "arm_compute/core/CL/kernels/CLGaussian5x5Kernel.h"
 #include "arm_compute/core/CL/kernels/CLGaussianPyramidKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLGatherKernel.h b/arm_compute/core/CL/kernels/CLGatherKernel.h
new file mode 100644
index 0000000000..4dac6b0d1f
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLGatherKernel.h
@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) 2018-2019 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_CLGATHERKERNEL_H__
+#define __ARM_COMPUTE_CLGATHERKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface for the kernel to perform tensor reshaping */
+class CLGatherKernel : public ICLKernel
+{
+public:
+    /** Default constructor */
+    CLGatherKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGatherKernel(const CLGatherKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGatherKernel &operator=(const CLGatherKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLGatherKernel(CLGatherKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLGatherKernel &operator=(CLGatherKernel &&) = default;
+    /** Default destructor */
+    ~CLGatherKernel() = default;
+    /** Initialise the kernel's inputs and outputs
+     *
+     * @param[in]  input   Source tensor. Supported tensor rank: up to 4. Data type supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32
+     * @param[in]  indices Indices tensor. Supported tensor rank: up to 1. Must be one of the following type: S64. Each value Must be in range [0, input.shape[@p axis])
+     * @param[out] output  Destination tensor. Data type supported: Same as @p input
+     * @param[in]  axis    (Optional) The axis in @p input to gather @p indices from. Negative values wrap around. Defaults to 0
+     */
+    void configure(const ICLTensor *input, const ICLTensor *indices, ICLTensor *output, int axis = 0);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref CLGatherKernel
+     *
+     * @param[in] input   Source tensor info. Supported tensor rank: up to 4. Data type supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32
+     * @param[in] indices Indices tensor info. Supported tensor rank: up to 4. Must be one of the following type: S64. Each value Must be in range [0, input.shape[@p axis])
+     * @param[in] output  Destination tensor info. Data type supported: Same as @p input
+     * @param[in] axis    (Optional) The axis in @p input to gather @p indices from. Negative values wrap around. Defaults to 0
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis = 0);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor *_input;   /**< Source tensor */
+    const ICLTensor *_indices; /**< Indices tensor */
+    ICLTensor       *_output;  /**< Destination tensor */
+    int              _axis;    /**< Axis index */
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGATHERKERNEL_H__ */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 4756ff4f97..ba0d8e254d 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -874,6 +874,18 @@ inline TensorShape compute_stack_shape(const ITensorInfo &a, unsigned int axis,
     }
     return shape_out;
 }
+
+inline TensorShape compute_gather_shape(const TensorShape &input_shape, const TensorShape &indices_shape, uint32_t actual_axis)
+{
+    ARM_COMPUTE_ERROR_ON(indices_shape.num_dimensions() > 1);
+    ARM_COMPUTE_ERROR_ON(input_shape.num_dimensions() > 4);
+    ARM_COMPUTE_ERROR_ON(actual_axis >= input_shape.num_dimensions());
+
+    TensorShape output_shape  = input_shape;
+    output_shape[actual_axis] = indices_shape[0];
+
+    return output_shape;
+}
 } // namespace shape_calculator
 } // namespace misc
 } // namespace arm_compute
diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index e0cf65b923..686d266557 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -76,6 +76,7 @@
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMTranspose1xW.h"
+#include "arm_compute/runtime/CL/functions/CLGather.h"
 #include "arm_compute/runtime/CL/functions/CLGaussian3x3.h"
 #include "arm_compute/runtime/CL/functions/CLGaussian5x5.h"
 #include "arm_compute/runtime/CL/functions/CLGaussianPyramid.h"
diff --git a/arm_compute/runtime/CL/functions/CLGather.h b/arm_compute/runtime/CL/functions/CLGather.h
new file mode 100644
index 0000000000..048804dfb2
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLGather.h
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2018-2019 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_CLGATHER_H__
+#define __ARM_COMPUTE_CLGATHER_H__
+
+#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Basic function to run @ref CLGatherKernel */
+class CLGather : public ICLSimpleFunction
+{
+public:
+    /** Initialise the kernel's inputs and outputs
+     *
+     * @param[in]  input   Source tensor. Supported tensor rank: up to 4. Data type supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32
+     * @param[in]  indices Indices tensor. Supported tensor rank: up to 1. Must be one of the following type: S64. Each value Must be in range [0, input.shape[@p axis])
+     * @param[out] output  Destination tensor. Data type supported: Same as @p input
+     * @param[in]  axis    (Optional) The axis in @p input to gather @p indices from. Defaults to 0
+     */
+    void configure(const ICLTensor *input, const ICLTensor *indices, ICLTensor *output, int axis = 0);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref CLGatherKernel
+     *
+     * @param[in] input   Source tensor info. Supported tensor rank: up to 4. Data type supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32
+     * @param[in] indices Indices tensor info. Supported tensor rank: up to 4. Must be one of the following types: S64. Each value Must be in range [0, input.shape[@p axis])
+     * @param[in] output  Destination tensor info. Data type supported: Same as @p input
+     * @param[in] axis    (Optional) The axis in @p input to gather @p indices from. Defaults to 0
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis = 0);
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_CLGATHER_H__ */
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 470a50f76b..2bc2d06827 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -262,6 +262,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "finalize", "optical_flow_pyramid_lk.cl" },
     { "fuse_batchnormalization_layer", "batchnormalization_layer.cl" },
     { "floor_layer", "floor.cl" },
+    { "gather", "gather.cl" },
     { "gaussian1x5_sub_x", "gaussian_pyramid.cl" },
     { "gaussian5x1_sub_y", "gaussian_pyramid.cl" },
     { "gemm_accumulate_biases", "gemm.cl" },
@@ -684,6 +685,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/floor.clembed"
     },
     {
+        "gather.cl",
+#include "./cl_kernels/gather.clembed"
+    },
+    {
         "gaussian_pyramid.cl",
 #include "./cl_kernels/gaussian_pyramid.clembed"
     },
diff --git a/src/core/CL/cl_kernels/gather.cl b/src/core/CL/cl_kernels/gather.cl
new file mode 100644
index 0000000000..34593ef60f
--- /dev/null
+++ b/src/core/CL/cl_kernels/gather.cl
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2018-2019 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 "helpers.h"
+
+#if defined(DATA_TYPE) && defined(AXIS)
+
+/** Performs the Gather operation along the chosen axis
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
+ * @note Axis should be given as a preprocessor argument using -DAXIS=axis. e.g. -DAXIS=1
+ * @attention Output tensor depth should be given as a preprocessor argument using -DOUTPUT_DIM_Z=size. e.g. -DOUTPUT_DIM_Z=16
+ * @attention Input tensor depth should be given as a preprocessor argument using -DINPUT_DIM_Z=size. e.g. -DINPUT_DIM_Z=16
+ *
+ *
+ * @param[in]  input_ptr                             Pointer to the source tensor. Supported data types: U8/S8/U16/S16/U32/S32/F16/F32
+ * @param[in]  input_stride_x                        Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  input_step_x                          input_stride_x * number of elements along X processed per work item (in bytes)
+ * @param[in]  input_stride_y                        Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_y                          input_stride_y * number of elements along Y processed per work item (in bytes)
+ * @param[in]  input_stride_z                        Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_z                          input_stride_z * number of elements along Z processed per work item (in bytes)
+ * @param[in]  input_stride_w                        Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_w                          input_stride_w * number of elements along W processed per work item (in bytes)
+ * @param[in]  input_offset_first_element_in_bytes   Offset of the first element in the source tensor
+ * @param[in]  indices_ptr                           Pointer to the indices vector. Supported data types: U32.
+ * @param[in]  indices_stride_x                      Stride of the indices vector in X dimension (in bytes)
+ * @param[in]  indices_step_x                        input_stride_x * number of elements along X processed per work item (in bytes)
+ * @param[in]  indices_offset_first_element_in_bytes Offset of the first element in the indices vector
+ * @param[out] output_ptr                            Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                       Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                         output_stride_x * number of elements along X processed per work item (in bytes)
+ * @param[in]  output_stride_y                       Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                         output_stride_y * number of elements along Y processed per work item (in bytes)
+ * @param[in]  output_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                         output_stride_z * number of elements along Z processed per work item (in bytes)
+ * @param[in]  output_stride_w                       Stride of the destination tensor in W dimension (in bytes)
+ * @param[in]  output_step_w                         output_stride_w * number of elements along W processed per work item (in bytes)
+ * @param[in]  output_offset_first_element_in_bytes  Offset of the first element in the destination tensor
+ */
+__kernel void gather(
+    TENSOR4D_DECLARATION(input),
+    VECTOR_DECLARATION(indices),
+    TENSOR4D_DECLARATION(output))
+{
+    const int px = get_global_id(0);
+    const int py = get_global_id(1);
+    const int pz = get_global_id(2) % OUTPUT_DIM_Z;
+    const int pw = get_global_id(2) / OUTPUT_DIM_Z;
+
+    const Tensor4D input   = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, INPUT_DIM_Z);
+    const Vector   indices = CONVERT_TO_VECTOR_STRUCT_NO_STEP(indices);
+    Tensor4D       output  = CONVERT_TO_TENSOR4D_STRUCT(output, OUTPUT_DIM_Z);
+
+#if AXIS == 0
+    const uint index                 = *(__global const uint *)vector_offset(&indices, px);
+    __global const uchar *input_addr = tensor4D_offset(&input, index, py, pz, pw);
+#elif AXIS == 1
+    const uint index                 = *(__global const uint *)vector_offset(&indices, py);
+    __global const uchar *input_addr = tensor4D_offset(&input, px, index, pz, pw);
+#elif AXIS == 2
+    const uint index                 = *(__global const uint *)vector_offset(&indices, pz);
+    __global const uchar *input_addr = tensor4D_offset(&input, px, py, index, pw);
+#elif AXIS == 3
+    const uint index                 = *(__global const uint *)vector_offset(&indices, pw);
+    __global const uchar *input_addr = tensor4D_offset(&input, px, py, pz, index);
+#endif //AXIS
+
+    *(__global DATA_TYPE *)output.ptr = *((__global const DATA_TYPE *)input_addr);
+}
+
+#endif //defined(DATA_TYPE) && defined(AXIS)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLGatherKernel.cpp b/src/core/CL/kernels/CLGatherKernel.cpp
new file mode 100644
index 0000000000..006e755b30
--- /dev/null
+++ b/src/core/CL/kernels/CLGatherKernel.cpp
@@ -0,0 +1,135 @@
+/*
+ * Copyright (c) 2018-2019 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/CLGatherKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include <string>
+
+namespace arm_compute
+{
+namespace
+{
+inline Status validate_arguments(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis)
+{
+    const uint32_t actual_axis = wrap_around(axis, static_cast<int>(input->num_dimensions()));
+    ARM_COMPUTE_RETURN_ERROR_ON(indices->num_dimensions() > 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+    ARM_COMPUTE_RETURN_ERROR_ON(actual_axis >= input->num_dimensions());
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16,
+                                                         DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
+        TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->tensor_shape(), indices->tensor_shape(), actual_axis);
+        ARM_COMPUTE_RETURN_ERROR_ON(output_shape.total_size() != output->tensor_shape().total_size());
+    }
+
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(indices, 1, DataType::U32);
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *indices, ITensorInfo *output, int axis)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, indices);
+    const uint32_t actual_axis = wrap_around(axis, static_cast<int>(input->num_dimensions()));
+    // Output auto initialization if not yet initialized
+    TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->tensor_shape(), indices->tensor_shape(), actual_axis);
+    auto_init_if_empty((*output), output_shape, 1, input->data_type());
+
+    // Create window
+    Window win = calculate_max_window(*output, Steps());
+    output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+
+    return std::make_pair(Status{}, win);
+}
+
+} // namespace
+
+CLGatherKernel::CLGatherKernel()
+    : _input(nullptr), _indices(nullptr), _output(nullptr), _axis(0)
+{
+}
+
+void CLGatherKernel::configure(const ICLTensor *input, const ICLTensor *indices, ICLTensor *output, int axis)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, indices);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), indices->info(), output->info(), axis));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), indices->info(), output->info(), axis);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+
+    _input   = input;
+    _output  = output;
+    _indices = indices;
+    _axis    = wrap_around(axis, static_cast<int>(input->info()->num_dimensions()));
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DOUTPUT_DIM_Z=" + support::cpp11::to_string(output->info()->dimension(2)));
+    build_opts.add_option("-DINPUT_DIM_Z=" + support::cpp11::to_string(input->info()->dimension(2)));
+    build_opts.add_option("-DAXIS=" + support::cpp11::to_string(_axis));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gather", build_opts.options()));
+    ICLKernel::configure_internal(win_config.second);
+}
+
+Status CLGatherKernel::validate(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, indices, output, axis));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), indices->clone().get(), output->clone().get(), axis).first);
+    return Status{};
+}
+
+void CLGatherKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    Window       window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+    unsigned int idx              = 0;
+    add_4D_tensor_argument(idx, _input, window_collapsed);
+    add_1D_tensor_argument(idx, _indices, window_collapsed);
+    add_4D_tensor_argument(idx, _output, window_collapsed);
+    enqueue(queue, *this, window_collapsed);
+}
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLGather.cpp b/src/runtime/CL/functions/CLGather.cpp
new file mode 100644
index 0000000000..459438e702
--- /dev/null
+++ b/src/runtime/CL/functions/CLGather.cpp
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2018-2019 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/runtime/CL/functions/CLGather.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/kernels/CLGatherKernel.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+void CLGather::configure(const ICLTensor *input, const ICLTensor *indices, ICLTensor *output, int axis)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLGatherKernel>();
+    k->configure(input, indices, output, axis);
+    _kernel = std::move(k);
+}
+
+Status CLGather::validate(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis)
+{
+    return CLGatherKernel::validate(input, indices, output, axis);
+}
+} // namespace arm_compute
diff --git a/tests/datasets/GatherDataset.h b/tests/datasets/GatherDataset.h
new file mode 100644
index 0000000000..29f2cccf57
--- /dev/null
+++ b/tests/datasets/GatherDataset.h
@@ -0,0 +1,169 @@
+/*
+ * Copyright (c) 2018-2019 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_TEST_GATHER_DATASET
+#define ARM_COMPUTE_TEST_GATHER_DATASET
+
+#include "utils/TypePrinter.h"
+
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace datasets
+{
+class GatherDataset
+{
+public:
+    using type = std::tuple<TensorShape, TensorShape, int>;
+
+    struct iterator
+    {
+        iterator(std::vector<TensorShape>::const_iterator input_shapes_it,
+                 std::vector<TensorShape>::const_iterator starts_values_it,
+                 std::vector<int>::const_iterator         axis_it)
+            : _input_shapes_it{ std::move(input_shapes_it) },
+              _indices_shapes_it{ std::move(starts_values_it) },
+              _axis_it{ std::move(axis_it) }
+        {
+        }
+
+        std::string description() const
+        {
+            std::stringstream description;
+            description << "InputShape=" << *_input_shapes_it << ":";
+            description << "IndicesShape=" << *_indices_shapes_it << ":";
+            description << "Axis=" << *_axis_it << ":";
+            return description.str();
+        }
+
+        GatherDataset::type operator*() const
+        {
+            return std::make_tuple(*_input_shapes_it, *_indices_shapes_it, *_axis_it);
+        }
+
+        iterator &operator++()
+        {
+            ++_input_shapes_it;
+            ++_indices_shapes_it;
+            ++_axis_it;
+            return *this;
+        }
+
+    private:
+        std::vector<TensorShape>::const_iterator _input_shapes_it;
+        std::vector<TensorShape>::const_iterator _indices_shapes_it;
+        std::vector<int>::const_iterator         _axis_it;
+    };
+
+    iterator begin() const
+    {
+        return iterator(_input_shapes.begin(), _indices_shapes.begin(), _axis.begin());
+    }
+
+    int size() const
+    {
+        return std::min(_input_shapes.size(), std::min(_indices_shapes.size(), _axis.size()));
+    }
+
+    void add_config(TensorShape input_shape, TensorShape indices_shape, int axis)
+    {
+        _input_shapes.emplace_back(std::move(input_shape));
+        _indices_shapes.emplace_back(std::move(indices_shape));
+        _axis.emplace_back(std::move(axis));
+    }
+
+protected:
+    GatherDataset()                 = default;
+    GatherDataset(GatherDataset &&) = default;
+
+private:
+    std::vector<TensorShape> _input_shapes{};
+    std::vector<TensorShape> _indices_shapes{};
+    std::vector<int>         _axis{};
+};
+
+class SmallGatherDataset final : public GatherDataset
+{
+public:
+    SmallGatherDataset()
+    {
+        // 2D input
+        add_config(TensorShape(15U, 15U), TensorShape(5U), 0);
+        add_config(TensorShape(15U, 15U), TensorShape(5U), 1);
+        add_config(TensorShape(5U, 5U), TensorShape(80U), -1);
+
+        // 3D input
+        add_config(TensorShape(5U, 5U, 5U), TensorShape(19U), 0);
+        add_config(TensorShape(5U, 4U, 6U), TensorShape(30U), 1);
+        add_config(TensorShape(3U, 5U, 7U), TensorShape(20U), 2);
+        add_config(TensorShape(5U, 4U, 6U), TensorShape(30U), -1);
+        add_config(TensorShape(3U, 5U, 7U), TensorShape(20U), -2);
+
+        // 4D input
+        add_config(TensorShape(4U, 3U, 4U, 5U), TensorShape(4U), 0);
+        add_config(TensorShape(4U, 3U, 5U, 5U), TensorShape(5U), 1);
+        add_config(TensorShape(4U, 3U, 2U, 5U), TensorShape(6U), 2);
+        add_config(TensorShape(3U, 4U, 4U, 6U), TensorShape(7U), 3);
+        add_config(TensorShape(4U, 3U, 5U, 5U), TensorShape(5U), -1);
+        add_config(TensorShape(4U, 3U, 2U, 5U), TensorShape(6U), -2);
+        add_config(TensorShape(3U, 4U, 4U, 6U), TensorShape(7U), -3);
+    }
+};
+
+class LargeGatherDataset final : public GatherDataset
+{
+public:
+    LargeGatherDataset()
+    {
+        // 2D input
+        add_config(TensorShape(150U, 150U), TensorShape(50U), 0);
+        add_config(TensorShape(150U, 150U), TensorShape(50U), 1);
+        add_config(TensorShape(150U, 150U), TensorShape(50U), -1);
+
+        // 3D input
+        add_config(TensorShape(50U, 40U, 60U), TensorShape(33U), 0);
+        add_config(TensorShape(40U, 50U, 60U), TensorShape(24U), 1);
+        add_config(TensorShape(70U, 80U, 100U), TensorShape(50U), 2);
+        add_config(TensorShape(40U, 50U, 60U), TensorShape(24U), -1);
+        add_config(TensorShape(70U, 80U, 100U), TensorShape(50U), -2);
+
+        // 4D input
+        add_config(TensorShape(30U, 40U, 20U, 20U), TensorShape(33U), 0);
+        add_config(TensorShape(23U, 10U, 60U, 20U), TensorShape(24U), 1);
+        add_config(TensorShape(14U, 20U, 10U, 31U), TensorShape(30U), 2);
+        add_config(TensorShape(34U, 10U, 40U, 20U), TensorShape(50U), 3);
+        add_config(TensorShape(23U, 10U, 60U, 20U), TensorShape(24U), -1);
+        add_config(TensorShape(14U, 20U, 10U, 31U), TensorShape(30U), -2);
+        add_config(TensorShape(34U, 10U, 40U, 20U), TensorShape(50U), -3);
+    }
+};
+
+} // namespace datasets
+} // namespace test
+} // namespace arm_compute
+
+#endif /* ARM_COMPUTE_TEST_GATHER_DATASET */
diff --git a/tests/validation/CL/Gather.cpp b/tests/validation/CL/Gather.cpp
new file mode 100644
index 0000000000..cc892a30d1
--- /dev/null
+++ b/tests/validation/CL/Gather.cpp
@@ -0,0 +1,210 @@
+/*
+ * Copyright (c) 2018-2019 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/Types.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "arm_compute/runtime/CL/functions/CLGather.h"
+
+#include "tests/CL/CLAccessor.h"
+#include "tests/datasets/GatherDataset.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/GatherFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+TEST_SUITE(CL)
+TEST_SUITE(Gather)
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
+        framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 27U), 1, DataType::F16),
+                                                TensorInfo(TensorShape(27U, 27U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(27U, 27U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(27U, 27U), 1, DataType::F32),     // Invalid Indices data type
+                                                TensorInfo(TensorShape(27U, 27U), 1, DataType::F32),     // Invalid Indices dimensionality
+                                                TensorInfo(TensorShape(5U, 5U, 5U, 5U, 5U), 1, DataType::F32),    // Invalid Input dimensionality
+                                                TensorInfo(TensorShape(27U, 27U), 1, DataType::F16),     // Mismatching data type input/output
+                                                TensorInfo(TensorShape(27U, 27U), 1, DataType::F32),     // Invalid positive axis value
+                                                TensorInfo(TensorShape(27U, 27U), 1, DataType::F16),     // Invalid negative axis value
+        }),
+        framework::dataset::make("IndicesInfo", {
+                                                TensorInfo(TensorShape(10U), 1, DataType::U32),
+                                                TensorInfo(TensorShape(10U), 1, DataType::U32),
+                                                TensorInfo(TensorShape(10U), 1, DataType::U32),
+                                                TensorInfo(TensorShape(10U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(10U, 10U), 1, DataType::U32),
+                                                TensorInfo(TensorShape(10U), 1, DataType::U32),
+                                                TensorInfo(TensorShape(10U), 1, DataType::U32),
+                                                TensorInfo(TensorShape(10U), 1, DataType::U32),
+                                                TensorInfo(TensorShape(10U), 1, DataType::U32),
+        })),
+        framework::dataset::make("OutputInfo", {
+                                                TensorInfo(TensorShape(10U, 27U), 1, DataType::F16),
+                                                TensorInfo(TensorShape(27U, 10U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(10U, 27U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(10U, 27U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(27U, 10U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(10U, 5U, 5U, 5U, 5U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(27U, 10U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(27U, 27U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(27U, 27U), 1, DataType::F16),
+        })),
+        framework::dataset::make("Axis", {
+                                            0,
+                                            1,
+                                            -2,
+                                            0,
+                                            1,
+                                            0,
+                                            1,
+                                            2,
+                                            -3,
+        })),
+        framework::dataset::make("Expected", { true, true, true, false, false, false, false, false, false })),
+        input_info, indices_info, output_info, axis, expected)
+{
+    const Status status = CLGather::validate(&input_info.clone()->set_is_resizable(true), &indices_info.clone()->set_is_resizable(true), &output_info.clone()->set_is_resizable(true), axis);
+    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+DATA_TEST_CASE(Configuration,
+               framework::DatasetMode::ALL,
+               combine(arm_compute::test::datasets::SmallGatherDataset(), framework::dataset::make("DataType", { DataType::F16, DataType::F32 })),
+               input_shape, indices_shape, axis, data_type)
+{
+    const uint32_t actual_axis = wrap_around(axis, static_cast<int>(input_shape.num_dimensions()));
+    CLTensor       src         = create_tensor<CLTensor>(input_shape, data_type);
+    CLTensor       indices     = create_tensor<CLTensor>(indices_shape, DataType::U32);
+    TensorShape    dst_shape   = arm_compute::misc::shape_calculator::compute_gather_shape(input_shape, indices_shape, actual_axis);
+    CLTensor       dst         = create_tensor<CLTensor>(dst_shape, data_type);
+
+    // Create and Configure function
+    CLGather gather;
+    gather.configure(&src, &indices, &dst, axis);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(dst.info()->tensor_shape());
+    validate(dst.info()->valid_region(), valid_region);
+}
+
+template <typename T>
+using CLGatherFixture = GatherFixture<CLTensor, CLAccessor, CLGather, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall,
+                       CLGatherFixture<half>,
+                       framework::DatasetMode::PRECOMMIT,
+                       combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+                       CLGatherFixture<half>,
+                       framework::DatasetMode::NIGHTLY,
+                       combine(datasets::LargeGatherDataset(), framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // FP16
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall,
+                       CLGatherFixture<float>,
+                       framework::DatasetMode::PRECOMMIT,
+                       combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+                       CLGatherFixture<float>,
+                       framework::DatasetMode::NIGHTLY,
+                       combine(datasets::LargeGatherDataset(), framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+
+TEST_SUITE(U8)
+FIXTURE_DATA_TEST_CASE(RunSmall,
+                       CLGatherFixture<uint8_t>,
+                       framework::DatasetMode::PRECOMMIT,
+                       combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::U8)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+                       CLGatherFixture<uint8_t>,
+                       framework::DatasetMode::NIGHTLY,
+                       combine(datasets::LargeGatherDataset(), framework::dataset::make("DataType", DataType::U8)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // U8
+
+TEST_SUITE(U16)
+FIXTURE_DATA_TEST_CASE(RunSmall,
+                       CLGatherFixture<uint16_t>,
+                       framework::DatasetMode::PRECOMMIT,
+                       combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::U16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+                       CLGatherFixture<uint16_t>,
+                       framework::DatasetMode::NIGHTLY,
+                       combine(datasets::LargeGatherDataset(), framework::dataset::make("DataType", DataType::U16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // U16
+
+TEST_SUITE_END() // Gather
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/GatherFixture.h b/tests/validation/fixtures/GatherFixture.h
new file mode 100644
index 0000000000..f2dcd4a1a4
--- /dev/null
+++ b/tests/validation/fixtures/GatherFixture.h
@@ -0,0 +1,142 @@
+/*
+ * Copyright (c) 2018-2019 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_TEST_GATHER_FIXTURE
+#define ARM_COMPUTE_TEST_GATHER_FIXTURE
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "tests/AssetsLibrary.h"
+#include "tests/Globals.h"
+#include "tests/IAccessor.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Fixture.h"
+#include "tests/validation/Helpers.h"
+#include "tests/validation/reference/Gather.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class GatherFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape indices_shape, int axis, DataType data_type)
+    {
+        _target    = compute_target(input_shape, data_type, axis, indices_shape);
+        _reference = compute_reference(input_shape, data_type, axis, indices_shape);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor)
+    {
+        library->fill_tensor_uniform(tensor, 0);
+    }
+
+    template <typename U>
+    void generate_indices(U &&indices, const TensorShape &input_shape, uint32_t actual_axis, TensorShape indices_shape)
+    {
+        std::mt19937 gen(library->seed());
+        uint32_t    *indices_ptr = static_cast<uint32_t *>(indices.data());
+
+        std::uniform_int_distribution<uint32_t> dist_index(0, input_shape[actual_axis] - 1);
+        //Let's consider 1D indices
+        for(unsigned int ind = 0; ind < indices_shape[0]; ind++)
+        {
+            indices_ptr[ind] = dist_index(gen);
+        }
+    }
+
+    TensorType compute_target(const TensorShape &input_shape,
+                              DataType           data_type,
+                              int                axis,
+                              const TensorShape  indices_shape)
+    {
+        // Create tensors
+        TensorType     src            = create_tensor<TensorType>(input_shape, data_type);
+        TensorType     indices_tensor = create_tensor<TensorType>(indices_shape, DataType::U32);
+        const uint32_t actual_axis    = wrap_around(axis, static_cast<int>(input_shape.num_dimensions()));
+        TensorShape    output_shape   = arm_compute::misc::shape_calculator::compute_gather_shape(input_shape, indices_shape, actual_axis);
+        TensorType     dst            = create_tensor<TensorType>(output_shape, data_type);
+
+        // Create and configure function
+        FunctionType gather;
+        gather.configure(&src, &indices_tensor, &dst, axis);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(indices_tensor.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        indices_tensor.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!indices_tensor.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src));
+        generate_indices(AccessorType(indices_tensor), input_shape, actual_axis, indices_shape);
+
+        // Compute function
+        gather.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape,
+                                      DataType           data_type,
+                                      int                axis,
+                                      const TensorShape  indices_shape)
+    {
+        // Create reference tensor
+        SimpleTensor<T>        src{ input_shape, data_type };
+        SimpleTensor<uint32_t> indices_tensor{ indices_shape, DataType::U32 };
+        const uint32_t         actual_axis = wrap_around(axis, static_cast<int>(input_shape.num_dimensions()));
+
+        // Fill reference tensor
+        fill(src);
+        generate_indices(indices_tensor, input_shape, actual_axis, indices_shape);
+
+        return reference::gather(src, indices_tensor, actual_axis);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+
+#endif /* ARM_COMPUTE_TEST_GATHER_FIXTURE */
diff --git a/tests/validation/reference/Gather.cpp b/tests/validation/reference/Gather.cpp
new file mode 100644
index 0000000000..ab5ea2f92e
--- /dev/null
+++ b/tests/validation/reference/Gather.cpp
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 2018-2019 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 "Gather.h"
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "tests/validation/Helpers.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+template <typename T>
+SimpleTensor<T> gather(const SimpleTensor<T> &src, const SimpleTensor<uint32_t> &indices, uint32_t actual_axis)
+{
+    const auto       *indices_ptr = static_cast<const uint32_t *>(indices.data());
+    const TensorShape dst_shape   = arm_compute::misc::shape_calculator::compute_gather_shape(src.shape(), indices.shape(), actual_axis);
+    SimpleTensor<T>   dst(dst_shape, src.data_type());
+
+    Window win;
+    win.use_tensor_dimensions(dst_shape);
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        Coordinates offset;
+        for(unsigned int dim = 0; dim < id.num_dimensions(); ++dim)
+        {
+            if(dim == actual_axis)
+            {
+                offset.set(dim, indices_ptr[id[dim]]);
+            }
+            else
+            {
+                offset.set(dim, id[dim]);
+            }
+        }
+        *reinterpret_cast<T *>(dst(id)) = *reinterpret_cast<const T *>(src(offset));
+    });
+
+    return dst;
+}
+
+template SimpleTensor<float> gather(const SimpleTensor<float> &src, const SimpleTensor<uint32_t> &indices, uint32_t actual_axis);
+template SimpleTensor<half> gather(const SimpleTensor<half> &src, const SimpleTensor<uint32_t> &indices, uint32_t actual_axis);
+template SimpleTensor<uint16_t> gather(const SimpleTensor<uint16_t> &src, const SimpleTensor<uint32_t> &indices, uint32_t actual_axis);
+template SimpleTensor<uint8_t> gather(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint32_t> &indices, uint32_t actual_axis);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/reference/Gather.h b/tests/validation/reference/Gather.h
new file mode 100644
index 0000000000..54e1cb8696
--- /dev/null
+++ b/tests/validation/reference/Gather.h
@@ -0,0 +1,46 @@
+/*
+ * Copyright (c) 2018-2019 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_TEST_GATHER_H__
+#define __ARM_COMPUTE_TEST_GATHER_H__
+
+#include "arm_compute/core/Types.h"
+#include "tests/SimpleTensor.h"
+#include "tests/validation/Helpers.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+template <typename T>
+SimpleTensor<T> gather(const SimpleTensor<T> &src, const SimpleTensor<uint32_t> &indices, uint32_t actual_axis);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+
+#endif /* __ARM_COMPUTE_TEST_GATHER_H__ */