COMPMID-1050 CL/NEON: Create a function to convert the 2D weights of FC layer from NHWC to NCHW and viceversa

Change-Id: If77ffeb92b6eb883e5d2d2c97c2c4d1d23d17c8d
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/129257
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

20 files changed, 1109 insertions, 4 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 2a9900d9fc..9316c6e489 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -42,6 +42,7 @@
 #include "arm_compute/core/CL/kernels/CLChannelShuffleLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLCol2ImKernel.h"
 #include "arm_compute/core/CL/kernels/CLColorConvertKernel.h"
+#include "arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h"
 #include "arm_compute/core/CL/kernels/CLConvolutionKernel.h"
 #include "arm_compute/core/CL/kernels/CLCopyKernel.h"
 #include "arm_compute/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h b/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h
new file mode 100644
index 0000000000..fe24aa9d8c
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h
@@ -0,0 +1,81 @@
+/*
+ * Copyright (c) 2018 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_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__
+#define __ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface to convert the 2D Fully Connected weights from NCHW to NHWC or vice versa.
+ *
+ * @note This function can be applied to the 2D weights used by a Fully Connected layer if:
+ *       - It follows a Convolution layer
+ *       - The data layout used by the network does not match the one the model has been trained in.
+ *
+ * @note This function assumes the weights are already reshaped (transposed)
+ */
+class CLConvertFullyConnectedWeightsKernel : public ICLKernel
+{
+public:
+    /** Default constructor */
+    CLConvertFullyConnectedWeightsKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLConvertFullyConnectedWeightsKernel(const CLConvertFullyConnectedWeightsKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLConvertFullyConnectedWeightsKernel &operator=(const CLConvertFullyConnectedWeightsKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLConvertFullyConnectedWeightsKernel(CLConvertFullyConnectedWeightsKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLConvertFullyConnectedWeightsKernel &operator=(CLConvertFullyConnectedWeightsKernel &&) = default;
+    /** Default destructor */
+    ~CLConvertFullyConnectedWeightsKernel() = default;
+    /** Set the input and output tensor.
+     *
+     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  data_layout          The data layout the weights have been trained in.
+     */
+    void configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLConvertFullyConnectedWeightsKernel
+     *
+     * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] data_layout          The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor *_input;
+    ICLTensor       *_output;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__ */
diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 0e271efa49..b9c54b2c25 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -41,6 +41,7 @@
 #include "arm_compute/core/NEON/kernels/NEChannelExtractKernel.h"
 #include "arm_compute/core/NEON/kernels/NECol2ImKernel.h"
 #include "arm_compute/core/NEON/kernels/NEColorConvertKernel.h"
+#include "arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h"
 #include "arm_compute/core/NEON/kernels/NEConvolutionKernel.h"
 #include "arm_compute/core/NEON/kernels/NECumulativeDistributionKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDepthConcatenateLayerKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h b/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h
new file mode 100644
index 0000000000..65ce764246
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h
@@ -0,0 +1,94 @@
+/*
+ * Copyright (c) 2018 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_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__
+#define __ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface to convert the 2D Fully Connected weights from NCHW to NHWC or vice versa.
+ *
+ * @note This function can be applied to the 2D weights used by a Fully Connected layer if:
+ *       - It follows a Convolution layer
+ *       - The data layout used by the network does not match the one the model has been trained in.
+ *
+ * @note This function assumes the weights are already reshaped (transposed)
+ */
+class NEConvertFullyConnectedWeightsKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEConvertFullyConnectedWeightsKernel";
+    }
+    /** Default constructor */
+    NEConvertFullyConnectedWeightsKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEConvertFullyConnectedWeightsKernel(const NEConvertFullyConnectedWeightsKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEConvertFullyConnectedWeightsKernel &operator=(const NEConvertFullyConnectedWeightsKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEConvertFullyConnectedWeightsKernel(NEConvertFullyConnectedWeightsKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEConvertFullyConnectedWeightsKernel &operator=(NEConvertFullyConnectedWeightsKernel &&) = default;
+    /** Default destructor */
+    ~NEConvertFullyConnectedWeightsKernel() = default;
+    /** Set the input and output tensor.
+     *
+     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  data_layout          The data layout the weights have been trained in.
+     */
+    void configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEConvertFullyConnectedWeightsKernel
+     *
+     * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] data_layout          The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    /** Template function to run the permute
+     *
+     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <typename T>
+    void run_convert_fc_weights(const Window &window);
+
+    const ITensor *_input;
+    ITensor       *_output;
+    unsigned int   _factor1; /*  equals to the number of elements per original input plane if @p data_layout == NCHW; its number of channels otherwise */
+    unsigned int   _factor2; /*  equals to the number of elements per original input plane if @p data_layout == NHWC; its number of channels otherwise */
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTSKERNEL_H__ */
diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index 53e49936df..0d9407bfe4 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -41,6 +41,7 @@
 #include "arm_compute/runtime/CL/functions/CLChannelExtract.h"
 #include "arm_compute/runtime/CL/functions/CLChannelShuffleLayer.h"
 #include "arm_compute/runtime/CL/functions/CLColorConvert.h"
+#include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h"
 #include "arm_compute/runtime/CL/functions/CLConvolution.h"
 #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h"
 #include "arm_compute/runtime/CL/functions/CLCopy.h"
diff --git a/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h b/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
new file mode 100644
index 0000000000..d8eef34c62
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2018 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_CLCONVERTFULLYCONNECTEDWEIGHTS_H__
+#define __ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTS_H__
+
+#include "arm_compute/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.h"
+#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Basic function to run @ref CLConvertFullyConnectedWeightsKernel. */
+class CLConvertFullyConnectedWeights : public ICLSimpleFunction
+{
+public:
+    /** Initialize the function.
+     *
+     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  data_layout          The data layout the weights have been trained in.
+     */
+    void configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLConvertFullyConnectedWeights
+     *
+     * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] data_layout          The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTS_H__ */
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 8091acd258..bd4097c6d3 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -41,6 +41,7 @@
 #include "arm_compute/runtime/NEON/functions/NEChannelExtract.h"
 #include "arm_compute/runtime/NEON/functions/NECol2Im.h"
 #include "arm_compute/runtime/NEON/functions/NEColorConvert.h"
+#include "arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h"
 #include "arm_compute/runtime/NEON/functions/NEConvolution.h"
 #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h"
@@ -111,4 +112,4 @@
 #include "arm_compute/runtime/NEON/functions/NEWarpPerspective.h"
 #include "arm_compute/runtime/NEON/functions/NEWinogradConvolutionLayer.h"
 
-#endif /* __ARM_COMPUTE_NEFUNCTIONS_H__ */
+#endif /* __ARM_COMPUTE_NEFUNCTIONS_H__ */
\ No newline at end of file
diff --git a/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h b/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h
new file mode 100644
index 0000000000..bdb157f30b
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h
@@ -0,0 +1,65 @@
+/*
+ * Copyright (c) 2018 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_NECONVERTFULLYCONNECTEDWEIGHTS_H__
+#define __ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTS_H__
+
+#include "arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h"
+#include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Basic function to run @ref NEConvertFullyConnectedWeightsKernel. */
+class NEConvertFullyConnectedWeights : public IFunction
+{
+public:
+    /** Default constructor */
+    NEConvertFullyConnectedWeights();
+    /** Initialize the function.
+     *
+     * @param[in]  input                Source weights tensor to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[out] output               The converted weights tensor. Shape and Data Type: Same as @p input.
+     * @param[in]  original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in]  data_layout          The data layout the weights have been trained in.
+     */
+    void configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEConvertFullyConnectedWeights
+     *
+     * @param[in] input                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/U32/S32/QS32/F16/F32.
+     * @param[in] output               The converted weights tensor info. Shape and Data Type: Same as @p input.
+     * @param[in] original_input_shape Shape of the original input tensor (the one entering fully connected layer). Must be in NCHW format.
+     * @param[in] data_layout          The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout);
+
+    // Inherited methods overriden:
+    void run() override;
+
+private:
+    NEConvertFullyConnectedWeightsKernel _kernel;
+};
+}
+#endif /* __ARM_COMPUTE_NECONVERTFULLYCONNECTEDWEIGHTS_H__ */
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index f1be935df3..220c7490f3 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -174,6 +174,9 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "concatenate_depth", "concatenate.cl" },
     { "convolution_rectangle", "convolution_rectangle.cl" },
     { "col2im", "col2im.cl" },
+    { "convert_depth_down", "depth_convert.cl" },
+    { "convert_depth_up", "depth_convert.cl" },
+    { "convert_fc_weights", "convert_fc_weights.cl" },
     { "convolution3x3_static", "convolution3x3.cl" },
     { "convolution5x5_static", "convolution5x5.cl" },
     { "convolution7x7_static", "convolution7x7.cl" },
@@ -184,8 +187,6 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "convolution_separable7x1_static", "convolution7x7.cl" },
     { "convolution_separable1x9_static", "convolution9x9.cl" },
     { "convolution_separable9x1_static", "convolution9x9.cl" },
-    { "convert_depth_down", "depth_convert.cl" },
-    { "convert_depth_up", "depth_convert.cl" },
     { "copy_tensor", "copy_tensor.cl" },
     { "copy_plane", "channel_extract.cl" },
     { "copy_planes_3p", "channel_combine.cl" },
@@ -434,6 +435,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/color_convert.clembed"
     },
     {
+        "convert_fc_weights.cl",
+#include "./cl_kernels/convert_fc_weights.clembed"
+    },
+    {
         "convolution3x3.cl",
 #include "./cl_kernels/convolution3x3.clembed"
     },
diff --git a/src/core/CL/cl_kernels/convert_fc_weights.cl b/src/core/CL/cl_kernels/convert_fc_weights.cl
new file mode 100644
index 0000000000..3c3e8b0dc4
--- /dev/null
+++ b/src/core/CL/cl_kernels/convert_fc_weights.cl
@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2018 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(FACTOR_1) && defined(FACTOR_2)
+/** Perform a NCHW -> NHWC or NHWC -> NCHW conversion for Fully Connected 2D weights.
+ *
+ * For NCHW -> NHWC, FACTOR_1 will be equal to the product of the first two dimensions of FullyConnectedLayer's input and FACTOR_2 will represent the number of channels of that tensor.
+ * For NHWC -> NCHW, FACTOR_1 and FACTOR_2 will hold the same values, but swapped.
+ *
+ * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
+ * @attention Original input tensor width*height and depth should be given as a preprocessor argument using -DFACTOR_1=size and -DFACTOR_2=size for NCHW and vice versa for NHWC. e.g. -DFACTOR_1=256 and -DFACTOR_2=128
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: U8, S8, QS8, QASYMM8, U16, S16, QS16, U32, S32, QS32, F16, F32
+ * @param[in]  src_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+__kernel void convert_fc_weights(
+    IMAGE_DECLARATION(src),
+    IMAGE_DECLARATION(dst))
+{
+    Image src = CONVERT_TO_IMAGE_STRUCT(src);
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(0) * dst_stride_x + (get_global_id(1) % FACTOR_1 * FACTOR_2 + get_global_id(1) / FACTOR_1) * dst_stride_y;
+
+    *((__global DATA_TYPE *)dst_addr) = *((__global DATA_TYPE *)src.ptr);
+}
+#endif // defined(DATA_TYPE) && defined(FACTOR_1) && defined(FACTOR_2)
diff --git a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp
new file mode 100644
index 0000000000..1b211b0adb
--- /dev/null
+++ b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2018 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/CLConvertFullyConnectedWeightsKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+CLConvertFullyConnectedWeightsKernel::CLConvertFullyConnectedWeightsKernel()
+    : _input(nullptr), _output(nullptr)
+{
+}
+
+void CLConvertFullyConnectedWeightsKernel::configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape,
+                                                     DataLayout data_layout)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(CLConvertFullyConnectedWeightsKernel::validate(input->info(), output->info(), original_input_shape, data_layout));
+
+    _input  = input;
+    _output = output;
+
+    const unsigned int num_elems_per_input_plane = original_input_shape.x() * original_input_shape.y();
+    const unsigned int num_channels              = original_input_shape.z();
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    if(data_layout == DataLayout::NCHW)
+    {
+        build_opts.add_option("-DFACTOR_1=" + support::cpp11::to_string(num_elems_per_input_plane));
+        build_opts.add_option("-DFACTOR_2=" + support::cpp11::to_string(num_channels));
+    }
+    else
+    {
+        build_opts.add_option("-DFACTOR_1=" + support::cpp11::to_string(num_channels));
+        build_opts.add_option("-DFACTOR_2=" + support::cpp11::to_string(num_elems_per_input_plane));
+    }
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("convert_fc_weights", build_opts.options()));
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+    ICLKernel::configure(win);
+}
+
+Status CLConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape,
+                                                      DataLayout data_layout)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32,
+                                                         DataType::QS32, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() != 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != original_input_shape.total_size_lower(3));
+    ARM_COMPUTE_RETURN_ERROR_ON(data_layout == DataLayout::UNKNOWN);
+
+    return Status{};
+}
+
+void CLConvertFullyConnectedWeightsKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    unsigned int idx = 0;
+    add_2D_tensor_argument(idx, _input, window);
+    add_2D_tensor_argument(idx, _output, window);
+    enqueue(queue, *this, window);
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
index c5d2528aa2..5c0a7351eb 100644
--- a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
@@ -25,7 +25,6 @@
 
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/Helpers.h"
diff --git a/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp b/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp
new file mode 100644
index 0000000000..b3746bddf2
--- /dev/null
+++ b/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp
@@ -0,0 +1,116 @@
+/*
+ * Copyright (c) 2018 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/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+
+using namespace arm_compute;
+
+NEConvertFullyConnectedWeightsKernel::NEConvertFullyConnectedWeightsKernel()
+    : _input(nullptr), _output(nullptr), _factor1(0), _factor2(0)
+{
+}
+
+void NEConvertFullyConnectedWeightsKernel::configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape,
+                                                     DataLayout data_layout)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(NEConvertFullyConnectedWeightsKernel::validate(input->info(), output->info(), original_input_shape, data_layout));
+
+    _input  = input;
+    _output = output;
+
+    const unsigned int num_elems_per_input_plane = original_input_shape.x() * original_input_shape.y();
+    const unsigned int num_channels              = original_input_shape.z();
+
+    // Set build options
+    if(data_layout == DataLayout::NCHW)
+    {
+        _factor1 = num_elems_per_input_plane;
+        _factor2 = num_channels;
+    }
+    else
+    {
+        _factor1 = num_channels;
+        _factor2 = num_elems_per_input_plane;
+    }
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+    INEKernel::configure(win);
+}
+
+Status NEConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape,
+                                                      DataLayout data_layout)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32,
+                                                         DataType::QS32, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() != 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != original_input_shape.total_size_lower(3));
+    ARM_COMPUTE_RETURN_ERROR_ON(data_layout == DataLayout::UNKNOWN);
+
+    return Status{};
+}
+
+template <typename T>
+void NEConvertFullyConnectedWeightsKernel::run_convert_fc_weights(const Window &window)
+{
+    const unsigned int dst_stride_x = _output->info()->strides_in_bytes().x();
+    const unsigned int dst_stride_y = _output->info()->strides_in_bytes().y();
+
+    Iterator input(_input, window);
+    Iterator output(_output, window);
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        *reinterpret_cast<T *>(output.ptr() + id.x() * dst_stride_x + (id.y() % _factor1 * _factor2 + id.y() / _factor1) * dst_stride_y) = *reinterpret_cast<T *>(input.ptr());
+    },
+    input);
+}
+
+void NEConvertFullyConnectedWeightsKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    switch(_input->info()->element_size())
+    {
+        case 1:
+            run_convert_fc_weights<uint8_t>(window);
+            break;
+        case 2:
+            run_convert_fc_weights<uint16_t>(window);
+            break;
+        case 4:
+            run_convert_fc_weights<uint32_t>(window);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Data type not supported.");
+            break;
+    }
+}
diff --git a/src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp b/src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp
new file mode 100644
index 0000000000..c226e56aff
--- /dev/null
+++ b/src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp
@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2018 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/CLConvertFullyConnectedWeights.h"
+
+using namespace arm_compute;
+
+void CLConvertFullyConnectedWeights::configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape,
+                                               DataLayout data_layout)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLConvertFullyConnectedWeightsKernel>();
+    k->configure(input, output, original_input_shape, data_layout);
+    _kernel = std::move(k);
+}
+
+Status CLConvertFullyConnectedWeights::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape,
+                                                DataLayout data_layout)
+{
+    return CLConvertFullyConnectedWeightsKernel::validate(input, output, original_input_shape, data_layout);
+}
\ No newline at end of file
diff --git a/src/runtime/NEON/functions/NEConvertFullyConnectedWeights.cpp b/src/runtime/NEON/functions/NEConvertFullyConnectedWeights.cpp
new file mode 100644
index 0000000000..b5b159ae31
--- /dev/null
+++ b/src/runtime/NEON/functions/NEConvertFullyConnectedWeights.cpp
@@ -0,0 +1,48 @@
+/*
+ * Copyright (c) 2018 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/NEON/functions/NEConvertFullyConnectedWeights.h"
+
+using namespace arm_compute;
+
+NEConvertFullyConnectedWeights::NEConvertFullyConnectedWeights()
+    : _kernel()
+{
+}
+
+void NEConvertFullyConnectedWeights::configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape,
+                                               DataLayout data_layout)
+{
+    _kernel.configure(input, output, original_input_shape, data_layout);
+}
+
+Status NEConvertFullyConnectedWeights::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape,
+                                                DataLayout data_layout)
+{
+    return NEConvertFullyConnectedWeightsKernel::validate(input, output, original_input_shape, data_layout);
+}
+
+void NEConvertFullyConnectedWeights::run()
+{
+    NEScheduler::get().schedule(&_kernel, Window::DimZ);
+}
diff --git a/tests/validation/CL/ConvertFullyConnectedWeights.cpp b/tests/validation/CL/ConvertFullyConnectedWeights.cpp
new file mode 100644
index 0000000000..f42f38bba5
--- /dev/null
+++ b/tests/validation/CL/ConvertFullyConnectedWeights.cpp
@@ -0,0 +1,102 @@
+/*
+ * Copyright (c) 2018 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/CLConvertFullyConnectedWeights.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/datasets/ShapeDatasets.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/ConvertFullyConnectedWeightsFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+auto params = combine(framework::dataset::make("WeightsWidth", { 16, 32, 64 }), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(ConvertFullyConnectedWeights)
+
+template <typename T>
+using CLConvertFullyConnectedWeightsFixture = ConvertFullyConnectedWeightsValidationFixture<CLTensor, CLAccessor, CLConvertFullyConnectedWeights, T>;
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLConvertFullyConnectedWeightsFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                    DataType::F32))))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLConvertFullyConnectedWeightsFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                        DataType::F32))))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLConvertFullyConnectedWeightsFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                   DataType::F16))))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLConvertFullyConnectedWeightsFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                       DataType::F16))))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLConvertFullyConnectedWeightsFixture<uint8_t>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                      DataType::QASYMM8))))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLConvertFullyConnectedWeightsFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), combine(params, framework::dataset::make("DataType",
+                       DataType::QASYMM8))))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/NEON/ConvertFullyConnectedWeights.cpp b/tests/validation/NEON/ConvertFullyConnectedWeights.cpp
new file mode 100644
index 0000000000..85316a3756
--- /dev/null
+++ b/tests/validation/NEON/ConvertFullyConnectedWeights.cpp
@@ -0,0 +1,104 @@
+/*
+ * Copyright (c) 2018 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/NEON/functions/NEConvertFullyConnectedWeights.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/datasets/ShapeDatasets.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/ConvertFullyConnectedWeightsFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+auto params = combine(framework::dataset::make("WeightsWidth", { 16, 32, 64 }), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(ConvertFullyConnectedWeights)
+
+template <typename T>
+using NEConvertFullyConnectedWeightsFixture = ConvertFullyConnectedWeightsValidationFixture<Tensor, Accessor, NEConvertFullyConnectedWeights, T>;
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEConvertFullyConnectedWeightsFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                    DataType::F32))))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEConvertFullyConnectedWeightsFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                        DataType::F32))))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEConvertFullyConnectedWeightsFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                   DataType::F16))))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEConvertFullyConnectedWeightsFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                       DataType::F16))))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END()
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEConvertFullyConnectedWeightsFixture<uint8_t>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), combine(params, framework::dataset::make("DataType",
+                                                                                                                      DataType::QASYMM8))))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEConvertFullyConnectedWeightsFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), combine(params, framework::dataset::make("DataType",
+                       DataType::QASYMM8))))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/ConvertFullyConnectedWeightsFixture.h b/tests/validation/fixtures/ConvertFullyConnectedWeightsFixture.h
new file mode 100644
index 0000000000..0fcef5c31f
--- /dev/null
+++ b/tests/validation/fixtures/ConvertFullyConnectedWeightsFixture.h
@@ -0,0 +1,127 @@
+/*
+ * Copyright (c) 2018 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_CONVERT_FULLY_CONNECTED_WEIGHTS_FIXTURE
+#define ARM_COMPUTE_TEST_CONVERT_FULLY_CONNECTED_WEIGHTS_FIXTURE
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.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/reference/ConvertFullyConnectedWeights.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ConvertFullyConnectedWeightsValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, unsigned int weights_w, DataLayout training_data_layout, DataType data_type)
+    {
+        const unsigned int height = input_shape.x() * input_shape.y() * input_shape.z();
+        const TensorShape  weights_shape(weights_w, height);
+
+        _target    = compute_target(input_shape, weights_shape, training_data_layout, data_type);
+        _reference = compute_reference(input_shape, weights_shape, training_data_layout, data_type);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        switch(tensor.data_type())
+        {
+            case DataType::QASYMM8:
+            {
+                std::uniform_int_distribution<uint8_t> distribution(0, 10);
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            case DataType::F32:
+            case DataType::F16:
+            {
+                std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            default:
+                library->fill_tensor_uniform(tensor, i);
+        }
+    }
+
+    TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const DataLayout training_data_layout, const DataType data_type)
+    {
+        // Create tensors
+        TensorType src = create_tensor<TensorType>(weights_shape, data_type);
+        TensorType dst = create_tensor<TensorType>(weights_shape, data_type);
+
+        // Create and configure function
+        FunctionType convert_weights;
+
+        convert_weights.configure(&src, &dst, input_shape, training_data_layout);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src), 0);
+
+        // Compute function
+        convert_weights.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const DataLayout training_data_layout, const DataType data_type)
+    {
+        // Create reference
+        SimpleTensor<T> src{ weights_shape, data_type };
+
+        // Fill reference
+        fill(src, 0);
+
+        return reference::convert_fully_connected_weights(src, input_shape, training_data_layout);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_CONVERT_FULLY_CONNECTED_WEIGHTS_FIXTURE */
diff --git a/tests/validation/reference/ConvertFullyConnectedWeights.cpp b/tests/validation/reference/ConvertFullyConnectedWeights.cpp
new file mode 100644
index 0000000000..b0f537fa0c
--- /dev/null
+++ b/tests/validation/reference/ConvertFullyConnectedWeights.cpp
@@ -0,0 +1,65 @@
+/*
+ * Copyright (c) 2018 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 "ConvertFullyConnectedWeights.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+template <typename T>
+SimpleTensor<T> convert_fully_connected_weights(const SimpleTensor<T> &src, const TensorShape &original_input_shape, const DataLayout training_data_layout)
+{
+    SimpleTensor<T> dst(src.shape(), src.data_type());
+
+    const bool         is_nchw_to_nhwc           = training_data_layout == DataLayout::NCHW;
+    const unsigned int num_elems_per_input_plane = original_input_shape.x() * original_input_shape.y();
+    const unsigned int num_channels              = original_input_shape.z();
+    const unsigned int factor_1                  = is_nchw_to_nhwc ? num_elems_per_input_plane : num_channels;
+    const unsigned int factor_2                  = is_nchw_to_nhwc ? num_channels : num_elems_per_input_plane;
+
+    for(int i = 0; i < src.num_elements(); ++i)
+    {
+        const Coordinates coords_in = index2coords(src.shape(), i);
+        const Coordinates coords_out(coords_in.x(), coords_in.y() % factor_1 * factor_2 + coords_in.y() / factor_1);
+
+        dst[coords2index(dst.shape(), coords_out)] = src[i];
+    }
+
+    return dst;
+}
+
+template SimpleTensor<uint8_t> convert_fully_connected_weights(const SimpleTensor<uint8_t> &src, const TensorShape &original_input_shape,
+                                                               const DataLayout training_data_layout);
+template SimpleTensor<half> convert_fully_connected_weights(const SimpleTensor<half> &src, const TensorShape &original_input_shape,
+                                                            const DataLayout training_data_layout);
+template SimpleTensor<float> convert_fully_connected_weights(const SimpleTensor<float> &src, const TensorShape &original_input_shape,
+                                                             const DataLayout training_data_layout);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/reference/ConvertFullyConnectedWeights.h b/tests/validation/reference/ConvertFullyConnectedWeights.h
new file mode 100644
index 0000000000..a9bbf13d53
--- /dev/null
+++ b/tests/validation/reference/ConvertFullyConnectedWeights.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2018 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_CONVERT_FULLY_CONNECTED_WEIGHTS_H__
+#define __ARM_COMPUTE_TEST_CONVERT_FULLY_CONNECTED_WEIGHTS_H__
+
+#include "tests/SimpleTensor.h"
+#include "tests/validation/Helpers.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+template <typename T>
+SimpleTensor<T> convert_fully_connected_weights(const SimpleTensor<T> &src, const TensorShape &original_input_shape, const DataLayout training_data_layout);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_TEST_CONVERT_FULLY_CONNECTED_WEIGHTS_H__ */