COMPMID-1959: Implements 2D FFT on OpenCL

Change-Id: I73cf3984a5463acc854c8a59dc2bd9a5234cd99c Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com> Reviewed-on: https://review.mlplatform.org/c/936 Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Tested-by: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
author: Georgios Pinitas <georgios.pinitas@arm.com> 2019-03-26 17:23:28 +0000
committer: Gian Marco Iodice <gianmarco.iodice@arm.com> 2019-04-11 09:34:26 +0000
commit: 8be9148814b88e5b0cabd5a4d2b1f4ff470a8c1c (patch)
tree: 760658b8c7b8917379467bd3fc119a5502faa850
parent: a50e702289af66944e860eafc7f3b32f6c5f30be (diff)
download: ComputeLibrary-8be9148814b88e5b0cabd5a4d2b1f4ff470a8c1c.tar.gz
38 files changed, 2962 insertions, 249 deletions
diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index e3ffcd0704..57498715c8 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -66,6 +66,7 @@
 #include "arm_compute/core/CL/kernels/CLErodeKernel.h"
 #include "arm_compute/core/CL/kernels/CLFFTDigitReverseKernel.h"
 #include "arm_compute/core/CL/kernels/CLFFTRadixStageKernel.h"
+#include "arm_compute/core/CL/kernels/CLFFTScaleKernel.h"
 #include "arm_compute/core/CL/kernels/CLFastCornersKernel.h"
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
 #include "arm_compute/core/CL/kernels/CLFlattenLayerKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLFFTDigitReverseKernel.h b/arm_compute/core/CL/kernels/CLFFTDigitReverseKernel.h
index 10652cdb4d..3082cb186f 100644
--- a/arm_compute/core/CL/kernels/CLFFTDigitReverseKernel.h
+++ b/arm_compute/core/CL/kernels/CLFFTDigitReverseKernel.h
@@ -26,6 +26,8 @@
 
 #include "arm_compute/core/CL/ICLKernel.h"
 
+#include "arm_compute/core/KernelDescriptors.h"
+
 namespace arm_compute
 {
 // Forward declarations
@@ -52,19 +54,19 @@ public:
      * @param[in]  input  Source tensor. Data types supported: F32.
      * @param[out] output Destination tensor. Data type supported: same as @p input
      * @param[in]  idx    Digit reverse index tensor. Data type supported: U32
-     * @param[in]  axis   Axis to perform digit reverse on.
+     * @param[in]  config Kernel configuration.
      */
-    void configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *idx, unsigned int axis);
+    void configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *idx, const FFTDigitReverseKernelInfo &config);
     /** Static function to check if given info will lead to a valid configuration of @ref CLFFTDigitReverseKernel
      *
      * @param[in] input  Source tensor info. Data types supported: F32.
      * @param[in] output Destination tensor info. Data type supported: same as @p input
      * @param[in] idx    Digit reverse index tensor info. Data type supported: U32
-     * @param[in] axis   Axis to perform digit reverse on.
+     * @param[in] config Kernel configuration.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *idx, unsigned int axis);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *idx, const FFTDigitReverseKernelInfo &config);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
diff --git a/arm_compute/core/CL/kernels/CLFFTRadixStageKernel.h b/arm_compute/core/CL/kernels/CLFFTRadixStageKernel.h
index 9de775eafa..16fa390e5d 100644
--- a/arm_compute/core/CL/kernels/CLFFTRadixStageKernel.h
+++ b/arm_compute/core/CL/kernels/CLFFTRadixStageKernel.h
@@ -59,7 +59,7 @@ public:
      * @param[out]    output Destination tensor. Can be nullptr. Data type supported: same as @p input
      * @param[in]     config FFT descriptor metadata.
      */
-    void configure(ICLTensor *input, ICLTensor *output, const FFTRadixStageKernelDescriptor &config);
+    void configure(ICLTensor *input, ICLTensor *output, const FFTRadixStageKernelInfo &config);
     /** Static function to check if given info will lead to a valid configuration of @ref CLFFTRadixStageKernel
      *
      * @param[in] input  Source tensor info. Data types supported: F32.
@@ -68,7 +68,7 @@ public:
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const FFTRadixStageKernelDescriptor &config);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const FFTRadixStageKernelInfo &config);
     /** Returns the radix that are support by the FFT kernel
      *
      * @return A set of supported radix
diff --git a/arm_compute/core/CL/kernels/CLFFTScaleKernel.h b/arm_compute/core/CL/kernels/CLFFTScaleKernel.h
new file mode 100644
index 0000000000..39ecac42af
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLFFTScaleKernel.h
@@ -0,0 +1,78 @@
+/*
+ * Copyright (c) 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_CLFFTSCALEKERNEL_H__
+#define __ARM_COMPUTE_CLFFTSCALEKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+#include "arm_compute/core/KernelDescriptors.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ICLTensor;
+
+/** Interface for the inverse fft scale kernel. */
+class CLFFTScaleKernel : public ICLKernel
+{
+public:
+    /** Constructor */
+    CLFFTScaleKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLFFTScaleKernel(const CLFFTScaleKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLFFTScaleKernel &operator=(const CLFFTScaleKernel &) = delete;
+    /** Default Move Constructor. */
+    CLFFTScaleKernel(CLFFTScaleKernel &&) = default;
+    /** Default move assignment operator */
+    CLFFTScaleKernel &operator=(CLFFTScaleKernel &&) = default;
+    /** Default destructor */
+    ~CLFFTScaleKernel() = default;
+    /** Set the input and output tensors.
+     *
+     * @param[in,out] input  Source tensor. Data types supported: F32.
+     * @param[out]    output Destination tensor. Data type supported: same as @p input
+     * @param[in]     config Kernel configuration
+     */
+    void configure(ICLTensor *input, ICLTensor *output, const FFTScaleKernelInfo &config);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLFFTScaleKernel
+     *
+     * @param[in] input  Source tensor info. Data types supported: F32.
+     * @param[in] output Destination tensor info. Data type supported: same as @p input
+     * @param[in] config Kernel configuration
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const FFTScaleKernelInfo &config);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    ICLTensor *_input;
+    ICLTensor *_output;
+    bool       _run_in_place;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLFFTSCALEKERNEL_H__ */
diff --git a/arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h b/arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h
index b835aa701b..804182b187 100644
--- a/arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h
+++ b/arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,11 +29,10 @@
 
 namespace arm_compute
 {
+// Forward declarations
 class ICLTensor;
 
-/** Interface for the pixelwise multiplication kernel.
- *
- */
+/** Interface for the pixelwise multiplication kernel. */
 class CLPixelWiseMultiplicationKernel : public ICLKernel
 {
 public:
@@ -83,5 +82,46 @@ private:
     const ICLTensor *_input2;
     ICLTensor       *_output;
 };
+
+/** Interface for the complex pixelwise multiplication kernel. */
+class CLComplexPixelWiseMultiplicationKernel : public ICLKernel
+{
+public:
+    /** Default constructor.*/
+    CLComplexPixelWiseMultiplicationKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLComplexPixelWiseMultiplicationKernel(const CLComplexPixelWiseMultiplicationKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLComplexPixelWiseMultiplicationKernel &operator=(const CLComplexPixelWiseMultiplicationKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLComplexPixelWiseMultiplicationKernel(CLComplexPixelWiseMultiplicationKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLComplexPixelWiseMultiplicationKernel &operator=(CLComplexPixelWiseMultiplicationKernel &&) = default;
+    /** Initialise the kernel's input, output and border mode.
+     *
+     * @param[in]  input1 An input tensor. Data types supported: F32. Number of channels supported: 2.
+     * @param[in]  input2 An input tensor. Data types supported: same as @p input1. Number of channels supported: same as @p input1.
+     * @param[out] output The output tensor, Data types supported: same as @p input1. Number of channels supported: same as @p input1.
+     */
+    void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLComplexPixelWiseMultiplicationKernel
+     *
+     * @param[in] input1 An input tensor info. Data types supported: F32. Number of channels supported: 2.
+     * @param[in] input2 An input tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1.
+     * @param[in] output The output tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
+
+private:
+    const ICLTensor *_input1;
+    const ICLTensor *_input2;
+    ICLTensor       *_output;
+};
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLPIXELWISEMULTIPLICATIONKERNEL_H__ */
diff --git a/arm_compute/core/KernelDescriptors.h b/arm_compute/core/KernelDescriptors.h
index 186dbfb6d8..83131f4296 100644
--- a/arm_compute/core/KernelDescriptors.h
+++ b/arm_compute/core/KernelDescriptors.h
@@ -26,10 +26,24 @@
 
 namespace arm_compute
 {
+/** Descriptor for FFT scale kernels */
+struct FFTScaleKernelInfo
+{
+    float scale{ 0.f };      /**< Axis to perform the kernel on. */
+    bool  conjugate{ true }; /**< Flag to conjugate the output/ */
+};
+
+/** Descriptor for FFT digit reverse kernels */
+struct FFTDigitReverseKernelInfo
+{
+    unsigned int axis{ 0 };          /**< Axis to perform the kernel on. */
+    bool         conjugate{ false }; /**< Flag to conjugate the output/ */
+};
+
 /** Descriptor used by the FFT core kernels */
-struct FFTRadixStageKernelDescriptor
+struct FFTRadixStageKernelInfo
 {
-    unsigned int axis{ 0 };               /**< Axis to run the FFT on. */
+    unsigned int axis{ 0 };               /**< Axis to run the kernel on. */
     unsigned int radix{ 0 };              /**< Radix to use. */
     unsigned int Nx{ 0 };                 /**< Nx coefficient. */
     bool         is_first_stage{ false }; /**< Flags if the FFT kernels is the first stage of a decomposed FFT. */
diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index f1021843a0..a4fcdc27ac 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -67,6 +67,8 @@
 #include "arm_compute/runtime/CL/functions/CLEqualizeHistogram.h"
 #include "arm_compute/runtime/CL/functions/CLErode.h"
 #include "arm_compute/runtime/CL/functions/CLFFT1D.h"
+#include "arm_compute/runtime/CL/functions/CLFFT2D.h"
+#include "arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h"
 #include "arm_compute/runtime/CL/functions/CLFastCorners.h"
 #include "arm_compute/runtime/CL/functions/CLFillBorder.h"
 #include "arm_compute/runtime/CL/functions/CLFlattenLayer.h"
diff --git a/arm_compute/runtime/CL/functions/CLFFT1D.h b/arm_compute/runtime/CL/functions/CLFFT1D.h
index 1612cf7f50..029023c524 100644
--- a/arm_compute/runtime/CL/functions/CLFFT1D.h
+++ b/arm_compute/runtime/CL/functions/CLFFT1D.h
@@ -28,6 +28,7 @@
 
 #include "arm_compute/core/CL/kernels/CLFFTDigitReverseKernel.h"
 #include "arm_compute/core/CL/kernels/CLFFTRadixStageKernel.h"
+#include "arm_compute/core/CL/kernels/CLFFTScaleKernel.h"
 #include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/FunctionDescriptors.h"
@@ -39,8 +40,9 @@ class ICLTensor;
 
 /** Basic function to execute one dimensional FFT. This function calls the following OpenCL kernels:
  *
- * -# @ref CLFFTDigitReverseKernel Performs digit reverse
- * -# @ref CLFFTRadixStageKernel   A list of FFT kernels depending on the radix decomposition
+ * -# @ref CLFFTDigitReverseKernel Performs digit reverse.
+ * -# @ref CLFFTRadixStageKernel   A list of FFT kernels depending on the radix decomposition.
+ * -# @ref CLFFTScaleKernel        Performs output scaling in case of in inverse FFT.
  */
 class CLFFT1D : public IFunction
 {
@@ -69,11 +71,13 @@ public:
 
 protected:
     CLMemoryGroup                            _memory_group;
-    CLTensor                                 _digit_reversed_input;
-    CLTensor                                 _digit_reverse_indices;
     CLFFTDigitReverseKernel                  _digit_reverse_kernel;
     std::unique_ptr<CLFFTRadixStageKernel[]> _fft_kernels;
+    CLFFTScaleKernel                         _scale_kernel;
+    CLTensor                                 _digit_reversed_input;
+    CLTensor                                 _digit_reverse_indices;
     unsigned int                             _num_ffts;
+    bool                                     _run_scale;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLFFT1D_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLFFT2D.h b/arm_compute/runtime/CL/functions/CLFFT2D.h
new file mode 100644
index 0000000000..a0673ecc96
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLFFT2D.h
@@ -0,0 +1,76 @@
+/*
+ * Copyright (c) 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_CLFFT2D_H__
+#define __ARM_COMPUTE_CLFFT2D_H__
+
+#include "arm_compute/runtime/IFunction.h"
+
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/functions/CLFFT1D.h"
+#include "arm_compute/runtime/FunctionDescriptors.h"
+
+namespace arm_compute
+{
+// Forward declaration
+class ICLTensor;
+
+/** Basic function to execute two dimensional FFT. This function calls the following OpenCL kernels:
+ *
+ * -# @ref CLFFT1D 1D FFT is performed on the first given axis
+ * -# @ref CLFFT1D 1D FFT is performed on the second given axis
+ */
+class CLFFT2D : public IFunction
+{
+public:
+    /** Default Constructor */
+    CLFFT2D(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Initialise the function's source, destinations and border mode.
+     *
+     * @param[in]  input  Source tensor. Data types supported: F32.
+     * @param[out] output Destination tensor. Data types and data layouts supported: Same as @p input.
+     * @param[in]  config FFT related configuration
+     */
+    void configure(const ICLTensor *input, ICLTensor *output, const FFT2DInfo &config);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLFFT2D.
+     *
+     * @param[in] input  Source tensor info. Data types supported: F32.
+     * @param[in] output Destination tensor info. Data types and data layouts supported: Same as @p input.
+     * @param[in] config FFT related configuration
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const FFT2DInfo &config);
+
+    // Inherited methods overridden:
+    void run() override;
+
+protected:
+    CLMemoryGroup _memory_group;
+    CLFFT1D       _first_pass_func;
+    CLFFT1D       _second_pass_func;
+    CLTensor      _first_pass_tensor;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLFFT2D_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h
new file mode 100644
index 0000000000..0fd2cf3cb1
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h
@@ -0,0 +1,154 @@
+/*
+ * Copyright (c) 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_CLFFTCONVOLUTIONLAYER_H__
+#define __ARM_COMPUTE_CLFFTCONVOLUTIONLAYER_H__
+
+#include "arm_compute/runtime/IFunction.h"
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/functions/CLActivationLayer.h"
+#include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h"
+#include "arm_compute/runtime/CL/functions/CLFFT2D.h"
+#include "arm_compute/runtime/CL/functions/CLPadLayer.h"
+#include "arm_compute/runtime/CL/functions/CLPermute.h"
+#include "arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h"
+#include "arm_compute/runtime/CL/functions/CLReductionOperation.h"
+#include "arm_compute/runtime/CL/functions/CLReshapeLayer.h"
+#include "arm_compute/runtime/CL/functions/CLReverse.h"
+#include "arm_compute/runtime/CL/functions/CLSlice.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ICLTensor;
+
+/** Basic function to execute FFT-based convolution on OpenCL. This function calls the following OpenCL functions/kernels:
+ *
+ *  -# @ref CLPermute                        Permute input if NHWC(only NCHW is supported).
+ *  -# @ref CLPadLayer                       Pad input.
+ *  -# @ref CLFFT2D                          Forward transform to the frequency domain.
+ *  -# @ref CLComplexPixelWiseMultiplication Complex element-wise product of input and the weights.
+ *  -# @ref CLReductionOperation             Reduction across channels.
+ *  -# @ref CLFFT2D                          Inverse transform back to the time domain.
+ *  -# @ref CLStridedSlice                   Extract valid output.
+ *  -# @ref CLArithmeticAddition             Add bias.
+ *  -# @ref CLActivationLayer                Perform activation.
+ *  -# @ref CLPermute                        Permute output if NHWC(only NCHW is supported).
+ */
+class CLFFTConvolutionLayer : public IFunction
+{
+public:
+    /** Default constructor */
+    CLFFTConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLFFTConvolutionLayer(const CLFFTConvolutionLayer &) = delete;
+    /** Default move constructor */
+    CLFFTConvolutionLayer(CLFFTConvolutionLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLFFTConvolutionLayer &operator=(const CLFFTConvolutionLayer &) = delete;
+    /** Default move assignment operator */
+    CLFFTConvolutionLayer &operator=(CLFFTConvolutionLayer &&) = default;
+    /** Set the input and output tensors.
+     *
+     * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
+     *
+     * @param[in]  input     Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                       while every optional dimension from 4 and above represent a batch of inputs.
+     *                       Data types supported: F32.
+     * @param[in]  weights   Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
+     * @param[in]  biases    Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input
+     * @param[out] output    Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                       Data types supported: Same as @p input.
+     * @param[in]  conv_info Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation.
+     */
+    void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                   const ActivationLayerInfo &act_info = ActivationLayerInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref CLFFTConvolutionLayer
+     *
+     * @note: This function only works with any square kernel size and unit strides for both NCHW and NHWC data layout
+     *
+     * @param[in]  input     Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                       while every optional dimension from 4 and above represent a batch of inputs.
+     *                       Data types supported: F32.
+     * @param[in]  weights   Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
+     * @param[in]  biases    Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p input
+     * @param[out] output    Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                       Data types supported: Same as @p input.
+     * @param[in]  conv_info Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  act_info  (Optional) Activation layer information in case of a fused activation.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                           const ActivationLayerInfo &act_info = ActivationLayerInfo());
+
+    // Inherited methods overridden:
+    void run() override;
+    void prepare() override;
+
+private:
+    CLMemoryGroup                    _memory_group;
+    CLReverse                        _flip_weights_func;
+    CLPermute                        _permute_input_func;
+    CLPermute                        _permute_output_func;
+    CLPermute                        _permute_weights_func;
+    CLPermute                        _permute_bias_func;
+    CLPadLayer                       _pad_input_func;
+    CLPadLayer                       _pad_weights_func;
+    CLFFT2D                          _transform_input_func;
+    CLFFT2D                          _transform_weights_func;
+    CLFFT2D                          _itransform_output_func;
+    CLComplexPixelWiseMultiplication _prod_func;
+    CLReductionOperation             _reduce_func;
+    CLSlice                          _extract_output_func;
+    CLArithmeticAddition             _bias_add_func;
+    CLActivationLayer                _activation_layer_func;
+
+    CLTensor _permuted_input;
+    CLTensor _permuted_weights;
+    CLTensor _permuted_bias;
+    CLTensor _permuted_output;
+    CLTensor _padded_input;
+    CLTensor _padded_weights;
+    CLTensor _flip_axis;
+    CLTensor _flipped_weights;
+    CLTensor _transformed_input;
+    CLTensor _transformed_weights;
+    CLTensor _input_weights_product;
+    CLTensor _output_product;
+    CLTensor _output_reduced;
+    CLTensor _itransformed_output;
+    CLTensor _reshaped_output;
+    CLTensor _bias_output;
+
+    const ICLTensor *_original_weights;
+    const ICLTensor *_original_bias;
+    bool             _is_activationlayer_enabled;
+    bool             _needs_permute;
+    bool             _has_bias;
+    bool             _is_prepared;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_CLFFTCONVOLUTIONLAYER_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h b/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h
index a59fb4aba8..0fa40a77f2 100644
--- a/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h
+++ b/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,6 +29,7 @@
 
 namespace arm_compute
 {
+// Forward declaration
 class ICLTensor;
 
 /** Basic function to run @ref CLPixelWiseMultiplicationKernel. */
@@ -64,5 +65,27 @@ public:
     static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale,
                            ConvertPolicy overflow_policy, RoundingPolicy rounding_policy);
 };
-}
+
+/** Basic function to run @ref CLComplexPixelWiseMultiplicationKernel. */
+class CLComplexPixelWiseMultiplication : public ICLSimpleFunction
+{
+public:
+    /** Initialise the kernel's inputs, output.
+     *
+     * @param[in, out] input1 An input tensor. Data types supported: F32. Number of channels supported: 2.
+     *                        The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
+     * @param[in, out] input2 An input tensor. Data types supported: same as @p input1. Number of channels supported: same as @p input1.
+     *                        The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
+     * @param[out]     output The output tensor, Data types supported: same as @p input1. Number of channels supported: same as @p input1.
+     */
+    void configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLComplexPixelWiseMultiplication
+     *
+     * @param[in] input1 An input tensor info. Data types supported: F32. Number of channels supported: 2.
+     * @param[in] input2 An input tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1.
+     * @param[in] output The output tensor info, Data types supported: same as @p input1. Number of channels supported: same as @p input1.
+     */
+    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output);
+};
+} // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLPIXELWISEMULTIPLICATION_H__ */
diff --git a/arm_compute/runtime/FunctionDescriptors.h b/arm_compute/runtime/FunctionDescriptors.h
index 7ff25019e6..f9b16e4218 100644
--- a/arm_compute/runtime/FunctionDescriptors.h
+++ b/arm_compute/runtime/FunctionDescriptors.h
@@ -24,12 +24,29 @@
 #ifndef __ARM_COMPUTE_RUNTIME_FUNCTION_DESCRIPTORS_H__
 #define __ARM_COMPUTE_RUNTIME_FUNCTION_DESCRIPTORS_H__
 
+#include <utility>
+
 namespace arm_compute
 {
-/** Descriptor used by the FFT1d function */
+/** FFT direction to use */
+enum class FFTDirection
+{
+    Forward,
+    Inverse
+};
+
+/** Descriptor used by the FFT1D function */
 struct FFT1DInfo
 {
-    unsigned int axis{ 0 }; /**< Axis to run the FFT on. */
+    unsigned int axis{ 0 };                          /**< Axis to run the FFT on. */
+    FFTDirection direction{ FFTDirection::Forward }; /**< Direction of the FFT. */
+};
+
+/** Descriptor used by the FFT2D function */
+struct FFT2DInfo
+{
+    std::pair<unsigned int, unsigned int> axes{ 0, 1 }; /**< Axes to run on. If same, multiple transforms are performed on single axis*/
+    FFTDirection direction{ FFTDirection::Forward };    /**< Direction of the FFT. */
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_RUNTIME_FUNCTION_DESCRIPTORS_H__ */
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 4fa8ac4142..322ff517d9 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -219,7 +219,6 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "depthwise_convolution_3x3_f16", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_nhwc", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_nhwc_stride1", "depthwise_convolution.cl" },
-    { "digit_reverse", "fft.cl" },
     { "dwc_3x3_native_qasymm8_nchw", "depthwise_convolution_quantized.cl" },
     { "dwc_3x3_native_qasymm8_dot8_nchw", "depthwise_convolution_quantized.cl" },
     { "dwc_3x3_reshaped_qasymm8_nhwc", "depthwise_convolution_quantized.cl" },
@@ -261,18 +260,33 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "elementwise_unary", "elementwise_unary.cl" },
     { "erode", "erode.cl" },
     { "fast_corners", "fast_corners.cl" },
+    { "fft_digit_reverse_axis_0", "fft_digit_reverse.cl" },
+    { "fft_digit_reverse_axis_1", "fft_digit_reverse.cl" },
     { "fft_radix_2_first_stage_axis_0", "fft.cl" },
+    { "fft_radix_2_first_stage_axis_1", "fft.cl" },
     { "fft_radix_2_axis_0", "fft.cl" },
+    { "fft_radix_2_axis_1", "fft.cl" },
     { "fft_radix_3_first_stage_axis_0", "fft.cl" },
+    { "fft_radix_3_first_stage_axis_1", "fft.cl" },
     { "fft_radix_3_axis_0", "fft.cl" },
+    { "fft_radix_3_axis_1", "fft.cl" },
     { "fft_radix_4_first_stage_axis_0", "fft.cl" },
+    { "fft_radix_4_first_stage_axis_1", "fft.cl" },
     { "fft_radix_4_axis_0", "fft.cl" },
+    { "fft_radix_4_axis_1", "fft.cl" },
     { "fft_radix_5_first_stage_axis_0", "fft.cl" },
+    { "fft_radix_5_first_stage_axis_1", "fft.cl" },
     { "fft_radix_5_axis_0", "fft.cl" },
+    { "fft_radix_5_axis_1", "fft.cl" },
     { "fft_radix_7_first_stage_axis_0", "fft.cl" },
+    { "fft_radix_7_first_stage_axis_1", "fft.cl" },
     { "fft_radix_7_axis_0", "fft.cl" },
+    { "fft_radix_7_axis_1", "fft.cl" },
     { "fft_radix_8_first_stage_axis_0", "fft.cl" },
+    { "fft_radix_8_first_stage_axis_1", "fft.cl" },
     { "fft_radix_8_axis_0", "fft.cl" },
+    { "fft_radix_8_axis_1", "fft.cl" },
+    { "fft_scale_conj", "fft_scale.cl" },
     { "fill_image_borders_constant", "fill_border.cl" },
     { "fill_image_borders_replicate", "fill_border.cl" },
     { "finalize", "optical_flow_pyramid_lk.cl" },
@@ -391,6 +405,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "NV21_to_YUV444_bt709", "color_convert.cl" },
     { "output_stage_quantized", "direct_convolution_1x1_3x3_5x5_quantized.cl" },
     { "permute", "permute.cl" },
+    { "pixelwise_mul_complex", "pixelwise_mul_float.cl" },
     { "pixelwise_mul_float", "pixelwise_mul_float.cl" },
     { "pixelwise_mul_int", "pixelwise_mul_int.cl" },
     { "pixelwise_mul_quantized", "pixelwise_mul_int.cl" },
@@ -710,6 +725,14 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/fft.clembed"
     },
     {
+        "fft_digit_reverse.cl",
+#include "./cl_kernels/fft_digit_reverse.clembed"
+    },
+    {
+        "fft_scale.cl",
+#include "./cl_kernels/fft_scale.clembed"
+    },
+    {
         "fill_border.cl",
 #include "./cl_kernels/fill_border.clembed"
     },
diff --git a/src/core/CL/cl_kernels/fft.cl b/src/core/CL/cl_kernels/fft.cl
index 5f1ef2483b..0027fd5b66 100644
--- a/src/core/CL/cl_kernels/fft.cl
+++ b/src/core/CL/cl_kernels/fft.cl
@@ -23,48 +23,6 @@
  */
 #include "helpers.h"
 
-/** Computes the digit reverse stage
- *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
- * @param[in]  src_stride_x                      Stride of the source tensor 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 tensor 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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in]  dst_stride_x                      Stride of the destination tensor 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 tensor 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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- * @param[in]  idx_ptr                           Pointer to the index tensor. Supported data types: U32
- * @param[in]  idx_stride_x                      Stride of the index tensor in X dimension (in bytes)
- * @param[in]  idx_step_x                        idx_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  idx_offset_first_element_in_bytes The offset of the first element in the index tensor
- */
-__kernel void digit_reverse(
-    TENSOR3D_DECLARATION(src),
-    TENSOR3D_DECLARATION(dst),
-    VECTOR_DECLARATION(idx))
-{
-    // Get tensor pointers
-    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(src);
-    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
-    Vector   idx = CONVERT_TO_VECTOR_STRUCT(idx);
-
-    const unsigned int iidx = *((__global uint *)(idx.ptr));
-
-    // Load data
-    float2 data = vload2(0, (__global float *)tensor3D_offset(&src, iidx, get_global_id(1), get_global_id(2)));
-
-    // Store result
-    vstore2(data, 0, (__global float *)dst.ptr);
-}
-
 /** Calculates and applies the twiddle factor to a given input.
  *
  * @param[in]     phi   The angle.
@@ -252,7 +210,7 @@ __kernel void digit_reverse(
         c7   = s4 + t1;                        \
     }
 
-/** Computes the first stage of a radix-2 DFT.
+/** Computes the first stage of a radix-2 DFT on axis 0.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -264,14 +222,14 @@ __kernel void digit_reverse(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  */
 kernel void fft_radix_2_first_stage_axis_0(
     TENSOR3D_DECLARATION(input)
@@ -289,17 +247,17 @@ kernel void fft_radix_2_first_stage_axis_0(
     Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
 #endif /* IN_PLACE */
 
-    // Load eight complex input values
+    // Load two complex input values
     float4 data = vload4(0, (__global float *)input.ptr);
 
     // Compute DFT N = 2
     DFT_2(data.s01, data.s23);
 
-    // Store eight complex output values
+    // Store two complex output values
     vstore4(data, 0, (__global float *)output.ptr);
 }
 
-/** Computes the first stage of a radix-3 DFT.
+/** Computes the first stage of a radix-2 DFT on axis 1.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -311,14 +269,63 @@ kernel void fft_radix_2_first_stage_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ */
+kernel void fft_radix_2_first_stage_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+)
+{
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+
+    // Load two complex input values
+    float2 data1 = vload2(0, (__global float *)input.ptr);
+    float2 data2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 1, 0));
+
+    // Compute DFT N = 2
+    DFT_2(data1, data2);
+
+    // Store two complex output values
+    vstore2(data1, 0, (__global float *)output.ptr);
+    vstore2(data2, 0, (__global float *)tensor3D_offset(&output, 0, 1, 0));
+}
+
+/** Computes the first stage of a radix-3 DFT on axis 0.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  */
 kernel void fft_radix_3_first_stage_axis_0(
     TENSOR3D_DECLARATION(input)
@@ -336,19 +343,19 @@ kernel void fft_radix_3_first_stage_axis_0(
     Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
 #endif /* IN_PLACE */
 
-    // Load eight complex input values
+    // Load three complex input values
     float4 data0 = vload4(0, (__global float *)input.ptr);
     float2 data1 = vload2(0, (__global float *)tensor3D_offset(&input, 2, 0, 0));
 
     // Compute DFT N = 3
     DFT_3(data0.s01, data0.s23, data1.s01);
 
-    // Store eight complex output values
+    // Store three complex output values
     vstore4(data0, 0, (__global float *)output.ptr);
     vstore2(data1, 0, (__global float *)tensor3D_offset(&output, 2, 0, 0));
 }
 
-/** Computes the first stage of a radix-4 DFT.
+/** Computes the first stage of a radix-3 DFT on axis 1.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -360,14 +367,65 @@ kernel void fft_radix_3_first_stage_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ */
+kernel void fft_radix_3_first_stage_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+)
+{
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+
+    // Load three complex input values
+    float2 data0 = vload2(0, (__global float *)input.ptr);
+    float2 data1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 1, 0));
+    float2 data2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2, 0));
+
+    // Compute DFT N = 3
+    DFT_3(data0, data1, data2);
+
+    // Store three complex output values
+    vstore2(data0, 0, (__global float *)output.ptr);
+    vstore2(data1, 0, (__global float *)tensor3D_offset(&output, 0, 1, 0));
+    vstore2(data2, 0, (__global float *)tensor3D_offset(&output, 0, 2, 0));
+}
+
+/** Computes the first stage of a radix-4 DFT on axis 0.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  */
 kernel void fft_radix_4_first_stage_axis_0(
     TENSOR3D_DECLARATION(input)
@@ -385,17 +443,70 @@ kernel void fft_radix_4_first_stage_axis_0(
     Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
 #endif /* IN_PLACE */
 
-    // Load eight complex input values
+    // Load four complex input values
     float8 data = vload8(0, (__global float *)input.ptr);
 
     // Compute DFT N = 4
     DFT_4(data.s01, data.s23, data.s45, data.s67);
 
-    // Store eight complex output values
+    // Store four complex output values
     vstore8(data, 0, (__global float *)output.ptr);
 }
 
-/** Computes the first stage of a radix-5 DFT.
+/** Computes the first stage of a radix-4 DFT on axis 1.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ */
+kernel void fft_radix_4_first_stage_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+)
+{
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+
+    // Load four complex input values
+    float2 data0 = vload2(0, (__global float *)input.ptr);
+    float2 data1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 1, 0));
+    float2 data2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2, 0));
+    float2 data3 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 3, 0));
+
+    // Compute DFT N = 4
+    DFT_4(data0, data1, data2, data3);
+
+    // Store four complex output values
+    vstore2(data0, 0, (__global float *)output.ptr);
+    vstore2(data1, 0, (__global float *)tensor3D_offset(&output, 0, 1, 0));
+    vstore2(data2, 0, (__global float *)tensor3D_offset(&output, 0, 2, 0));
+    vstore2(data3, 0, (__global float *)tensor3D_offset(&output, 0, 3, 0));
+}
+
+/** Computes the first stage of a radix-5 DFT on axis 0.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -407,14 +518,14 @@ kernel void fft_radix_4_first_stage_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  */
 kernel void fft_radix_5_first_stage_axis_0(
     TENSOR3D_DECLARATION(input)
@@ -432,19 +543,19 @@ kernel void fft_radix_5_first_stage_axis_0(
     Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
 #endif /* IN_PLACE */
 
-    // Load eight complex input values
+    // Load five complex input values
     float8 data0 = vload8(0, (__global float *)input.ptr);
     float2 data1 = vload2(0, (__global float *)tensor3D_offset(&input, 4, 0, 0));
 
     // Compute DFT N = 5
     DFT_5(data0.s01, data0.s23, data0.s45, data0.s67, data1.s01);
 
-    // Store eight complex output values
+    // Store five complex output values
     vstore8(data0, 0, (__global float *)output.ptr);
     vstore2(data1, 0, (__global float *)tensor3D_offset(&output, 4, 0, 0));
 }
 
-/** Computes the first stage of a radix-7 DFT.
+/** Computes the first stage of a radix-5 DFT on axis 1.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -456,14 +567,69 @@ kernel void fft_radix_5_first_stage_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ */
+kernel void fft_radix_5_first_stage_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+)
+{
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+
+    // Load five complex input values
+    float2 data0 = vload2(0, (__global float *)input.ptr);
+    float2 data1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 1, 0));
+    float2 data2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2, 0));
+    float2 data3 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 3, 0));
+    float2 data4 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 4, 0));
+
+    // Compute DFT N = 5
+    DFT_5(data0, data1, data2, data3, data4);
+
+    // Store five complex output values
+    vstore2(data0, 0, (__global float *)output.ptr);
+    vstore2(data1, 0, (__global float *)tensor3D_offset(&output, 0, 1, 0));
+    vstore2(data2, 0, (__global float *)tensor3D_offset(&output, 0, 2, 0));
+    vstore2(data3, 0, (__global float *)tensor3D_offset(&output, 0, 3, 0));
+    vstore2(data4, 0, (__global float *)tensor3D_offset(&output, 0, 4, 0));
+}
+
+/** Computes the first stage of a radix-7 DFT on axis 0.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  */
 kernel void fft_radix_7_first_stage_axis_0(
     TENSOR3D_DECLARATION(input)
@@ -481,7 +647,7 @@ kernel void fft_radix_7_first_stage_axis_0(
     Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
 #endif /* IN_PLACE */
 
-    // Load eight complex input values
+    // Load seven complex input values
     float8 data0 = vload8(0, (__global float *)input.ptr);
     float4 data1 = vload4(0, (__global float *)tensor3D_offset(&input, 4, 0, 0));
     float2 data2 = vload2(0, (__global float *)tensor3D_offset(&input, 6, 0, 0));
@@ -489,13 +655,72 @@ kernel void fft_radix_7_first_stage_axis_0(
     // Compute DFT N = 7
     DFT_7(data0.s01, data0.s23, data0.s45, data0.s67, data1.s01, data1.s23, data2.s01);
 
-    // Store eight complex output values
+    // Store seven complex output values
     vstore8(data0, 0, (__global float *)output.ptr);
     vstore4(data1, 0, (__global float *)tensor3D_offset(&output, 4, 0, 0));
     vstore2(data2, 0, (__global float *)tensor3D_offset(&output, 6, 0, 0));
 }
 
-/** Computes the first stage of a radix-8 DFT.
+/** Computes the first stage of a radix-7 DFT on axis 1.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ */
+kernel void fft_radix_7_first_stage_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+)
+{
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+
+    // Load seven complex input values
+    float2 data0 = vload2(0, (__global float *)input.ptr);
+    float2 data1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 1, 0));
+    float2 data2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2, 0));
+    float2 data3 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 3, 0));
+    float2 data4 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 4, 0));
+    float2 data5 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 5, 0));
+    float2 data6 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 6, 0));
+
+    // Compute DFT N = 7
+    DFT_7(data0, data1, data2, data3, data4, data5, data6);
+
+    // Store seven complex output values
+    vstore2(data0, 0, (__global float *)output.ptr);
+    vstore2(data1, 0, (__global float *)tensor3D_offset(&output, 0, 1, 0));
+    vstore2(data2, 0, (__global float *)tensor3D_offset(&output, 0, 2, 0));
+    vstore2(data3, 0, (__global float *)tensor3D_offset(&output, 0, 3, 0));
+    vstore2(data4, 0, (__global float *)tensor3D_offset(&output, 0, 4, 0));
+    vstore2(data5, 0, (__global float *)tensor3D_offset(&output, 0, 5, 0));
+    vstore2(data6, 0, (__global float *)tensor3D_offset(&output, 0, 6, 0));
+}
+
+/** Computes the first stage of a radix-8 DFT on axis 0.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -507,14 +732,14 @@ kernel void fft_radix_7_first_stage_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  */
 kernel void fft_radix_8_first_stage_axis_0(
     TENSOR3D_DECLARATION(input)
@@ -542,7 +767,68 @@ kernel void fft_radix_8_first_stage_axis_0(
     vstore16(data, 0, (__global float *)output.ptr);
 }
 
-/** Computes a stage of a radix-2 FFT.
+/** Computes the first stage of a radix-8 DFT on axis 1.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ */
+kernel void fft_radix_8_first_stage_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+)
+{
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+
+    // Load eight complex input values
+    float2 data0 = vload2(0, (__global float *)input.ptr);
+    float2 data1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 1, 0));
+    float2 data2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2, 0));
+    float2 data3 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 3, 0));
+    float2 data4 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 4, 0));
+    float2 data5 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 5, 0));
+    float2 data6 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 6, 0));
+    float2 data7 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 7, 0));
+
+    // Compute DFT N = 8
+    DFT_8(data0, data1, data2, data3, data4, data5, data6, data7);
+
+    // Store eight complex output values
+    vstore2(data0, 0, (__global float *)output.ptr);
+    vstore2(data1, 0, (__global float *)tensor3D_offset(&output, 0, 1, 0));
+    vstore2(data2, 0, (__global float *)tensor3D_offset(&output, 0, 2, 0));
+    vstore2(data3, 0, (__global float *)tensor3D_offset(&output, 0, 3, 0));
+    vstore2(data4, 0, (__global float *)tensor3D_offset(&output, 0, 4, 0));
+    vstore2(data5, 0, (__global float *)tensor3D_offset(&output, 0, 5, 0));
+    vstore2(data6, 0, (__global float *)tensor3D_offset(&output, 0, 6, 0));
+    vstore2(data7, 0, (__global float *)tensor3D_offset(&output, 0, 7, 0));
+}
+
+/** Computes a stage of a radix-2 FFT on axis 0.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -554,14 +840,14 @@ kernel void fft_radix_8_first_stage_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
  * @param[in]     Ni                                   Nx * Ny.
  * @param[in]     exp_const                            Exponent constant
@@ -612,7 +898,77 @@ kernel void fft_radix_2_axis_0(
     vstore2(c1, 0, (__global float *)tensor3D_offset(&output, Nx, 0, 0));
 }
 
-/** Computes a stage of a radix-3 FFT.
+/** Computes a stage of a radix-2 FFT on axis 1.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
+ * @param[in]     Ni                                   Nx * Ny.
+ * @param[in]     exp_const                            Exponent constant
+ */
+kernel void fft_radix_2_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+    ,
+    uint Nx, uint Ni, float exp_const)
+{
+    // Each work-item computes a single radix-2
+    uint kx = get_global_id(1);
+
+    // Compute nx
+    uint nx = kx % Nx;
+
+    // Compute n index
+    uint n = nx + (kx / Nx) * Ni;
+
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(input);
+    input.ptr += get_global_id(0) * input.stride_x + n * input.stride_y + get_global_id(2) * input.stride_z;
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
+    output.ptr += get_global_id(0) * output.stride_x + n * output.stride_y + get_global_id(2) * output.stride_z;
+#endif /* IN_PLACE */
+
+    // Load two complex input values
+    float2 c0 = vload2(0, (__global float *)input.ptr);
+    float2 c1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, Nx, 0));
+
+    // Compute phi
+    float phi = (float)nx * exp_const;
+
+    // Multiply by twiddle factor
+    TWIDDLE_FACTOR_MULTIPLICATION(phi, c1);
+
+    // Compute DFT N = 2
+    DFT_2(c0, c1);
+
+    // Store two complex output values
+    vstore2(c0, 0, (__global float *)output.ptr);
+    vstore2(c1, 0, (__global float *)tensor3D_offset(&output, 0, Nx, 0));
+}
+
+/** Computes a stage of a radix-3 FFT on axis 0.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -624,14 +980,14 @@ kernel void fft_radix_2_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
  * @param[in]     Ni                                   Nx * Ny.
  * @param[in]     exp_const                            Exponent constant
@@ -685,7 +1041,80 @@ kernel void fft_radix_3_axis_0(
     vstore2(c2, 0, (__global float *)tensor3D_offset(&output, 2 * Nx, 0, 0));
 }
 
-/** Computes a stage of a radix-4 FFT.
+/** Computes a stage of a radix-3 FFT on axis 1.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
+ * @param[in]     Ni                                   Nx * Ny.
+ * @param[in]     exp_const                            Exponent constant
+ */
+kernel void fft_radix_3_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+    ,
+    uint Nx, uint Ni, float exp_const)
+{
+    // Each work-item computes a single radix-3
+    uint kx = get_global_id(1);
+
+    // Compute nx
+    uint nx = kx % Nx;
+
+    // Compute n index
+    uint n = nx + (kx / Nx) * Ni;
+
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(input);
+    input.ptr += get_global_id(0) * input.stride_x + n * input.stride_y + get_global_id(2) * input.stride_z;
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
+    output.ptr += get_global_id(0) * output.stride_x + n * output.stride_y + get_global_id(2) * output.stride_z;
+#endif /* IN_PLACE */
+
+    // Load three complex input values
+    float2 c0 = vload2(0, (__global float *)input.ptr);
+    float2 c1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, Nx, 0));
+    float2 c2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2 * Nx, 0));
+
+    // Compute phi
+    float phi = (float)nx * exp_const;
+
+    // Multiply by twiddle factor
+    TWIDDLE_FACTOR_MULTIPLICATION(phi, c1);
+    TWIDDLE_FACTOR_MULTIPLICATION(2 * phi, c2);
+
+    // Compute DFT N = 3
+    DFT_3(c0, c1, c2);
+
+    // Store three complex output values
+    vstore2(c0, 0, (__global float *)output.ptr);
+    vstore2(c1, 0, (__global float *)tensor3D_offset(&output, 0, Nx, 0));
+    vstore2(c2, 0, (__global float *)tensor3D_offset(&output, 0, 2 * Nx, 0));
+}
+
+/** Computes a stage of a radix-4 FFT on axis 0.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -697,14 +1126,14 @@ kernel void fft_radix_3_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
  * @param[in]     Ni                                   Nx * Ny.
  * @param[in]     exp_const                            Exponent constant
@@ -761,7 +1190,7 @@ kernel void fft_radix_4_axis_0(
     vstore2(c3, 0, (__global float *)tensor3D_offset(&output, 3 * Nx, 0, 0));
 }
 
-/** Computes a stage of a radix-5 FFT.
+/** Computes a stage of a radix-4 FFT on axis 1.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -773,14 +1202,90 @@ kernel void fft_radix_4_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
+ * @param[in]     Ni                                   Nx * Ny.
+ * @param[in]     exp_const                            Exponent constant
+ */
+kernel void fft_radix_4_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+    ,
+    uint Nx, uint Ni, float exp_const)
+{
+    // Each work-item computes a single radix-4
+    uint kx = get_global_id(1);
+
+    // Compute nx
+    uint nx = kx % Nx;
+
+    // Compute n index
+    uint n = nx + (kx / Nx) * Ni;
+
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(input);
+    input.ptr += get_global_id(0) * input.stride_x + n * input.stride_y + get_global_id(2) * input.stride_z;
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
+    output.ptr += get_global_id(0) * output.stride_x + n * output.stride_y + get_global_id(2) * output.stride_z;
+#endif /* IN_PLACE */
+
+    // Load four complex input values
+    float2 c0 = vload2(0, (__global float *)input.ptr);
+    float2 c1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, Nx, 0));
+    float2 c2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2 * Nx, 0));
+    float2 c3 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 3 * Nx, 0));
+
+    // Compute phi
+    float phi = (float)nx * exp_const;
+
+    // Multiply by twiddle factor
+    TWIDDLE_FACTOR_MULTIPLICATION(phi, c1);
+    TWIDDLE_FACTOR_MULTIPLICATION(2 * phi, c2);
+    TWIDDLE_FACTOR_MULTIPLICATION(3 * phi, c3);
+
+    // Compute DFT N = 4
+    DFT_4(c0, c1, c2, c3);
+
+    // Store four complex output values
+    vstore2(c0, 0, (__global float *)output.ptr);
+    vstore2(c1, 0, (__global float *)tensor3D_offset(&output, 0, Nx, 0));
+    vstore2(c2, 0, (__global float *)tensor3D_offset(&output, 0, 2 * Nx, 0));
+    vstore2(c3, 0, (__global float *)tensor3D_offset(&output, 0, 3 * Nx, 0));
+}
+
+/** Computes a stage of a radix-5 FFT on axis 0.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
  * @param[in]     Ni                                   Nx * Ny.
  * @param[in]     exp_const                            Exponent constant
@@ -840,7 +1345,7 @@ kernel void fft_radix_5_axis_0(
     vstore2(c4, 0, (__global float *)tensor3D_offset(&output, 4 * Nx, 0, 0));
 }
 
-/** Computes a stage of a radix-7 FFT.
+/** Computes a stage of a radix-5 FFT on axis 1.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -852,14 +1357,93 @@ kernel void fft_radix_5_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
+ * @param[in]     Ni                                   Nx * Ny.
+ * @param[in]     exp_const                            Exponent constant
+ */
+kernel void fft_radix_5_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+    ,
+    uint Nx, uint Ni, float exp_const)
+{
+    // Each work-item computes a single radix-5
+    uint kx = get_global_id(1);
+
+    // Compute nx
+    uint nx = kx % Nx;
+
+    // Compute n index
+    uint n = nx + (kx / Nx) * Ni;
+
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(input);
+    input.ptr += get_global_id(0) * input.stride_x + n * input.stride_y + get_global_id(2) * input.stride_z;
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
+    output.ptr += get_global_id(0) * output.stride_x + n * output.stride_y + get_global_id(2) * output.stride_z;
+#endif /* IN_PLACE */
+
+    // Load five complex input values
+    float2 c0 = vload2(0, (__global float *)input.ptr);
+    float2 c1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, Nx, 0));
+    float2 c2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2 * Nx, 0));
+    float2 c3 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 3 * Nx, 0));
+    float2 c4 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 4 * Nx, 0));
+
+    // Compute phi
+    float phi = (float)nx * exp_const;
+
+    // Multiply by twiddle factor
+    TWIDDLE_FACTOR_MULTIPLICATION(phi, c1);
+    TWIDDLE_FACTOR_MULTIPLICATION(2 * phi, c2);
+    TWIDDLE_FACTOR_MULTIPLICATION(3 * phi, c3);
+    TWIDDLE_FACTOR_MULTIPLICATION(4 * phi, c4);
+
+    // Compute DFT N = 5
+    DFT_5(c0, c1, c2, c3, c4);
+
+    // Store five complex output values
+    vstore2(c0, 0, (__global float *)output.ptr);
+    vstore2(c1, 0, (__global float *)tensor3D_offset(&output, 0, Nx, 0));
+    vstore2(c2, 0, (__global float *)tensor3D_offset(&output, 0, 2 * Nx, 0));
+    vstore2(c3, 0, (__global float *)tensor3D_offset(&output, 0, 3 * Nx, 0));
+    vstore2(c4, 0, (__global float *)tensor3D_offset(&output, 0, 4 * Nx, 0));
+}
+
+/** Computes a stage of a radix-7 FFT on axis 0.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
  * @param[in]     Ni                                   Nx * Ny.
  * @param[in]     exp_const                            Exponent constant
@@ -925,7 +1509,92 @@ kernel void fft_radix_7_axis_0(
     vstore2(c6, 0, (__global float *)tensor3D_offset(&output, 6 * Nx, 0, 0));
 }
 
-/** Computes a stage of a radix-8 FFT.
+/** Computes a stage of a radix-7 FFT on axis 1.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
+ * @param[in]     Ni                                   Nx * Ny.
+ * @param[in]     exp_const                            Exponent constant
+ */
+kernel void fft_radix_7_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+    ,
+    uint Nx, uint Ni, float exp_const)
+{
+    // Each work-item computes a single radix-7
+    uint kx = get_global_id(1);
+
+    // Compute nx
+    uint nx = kx % Nx;
+
+    // Compute n index
+    uint n = nx + (kx / Nx) * Ni;
+
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(input);
+    input.ptr += get_global_id(0) * input.stride_x + n * input.stride_y + get_global_id(2) * input.stride_z;
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
+    output.ptr += get_global_id(0) * output.stride_x + n * output.stride_y + get_global_id(2) * output.stride_z;
+#endif /* IN_PLACE */
+
+    // Load seven complex input values
+    float2 c0 = vload2(0, (__global float *)input.ptr);
+    float2 c1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, Nx, 0));
+    float2 c2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2 * Nx, 0));
+    float2 c3 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 3 * Nx, 0));
+    float2 c4 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 4 * Nx, 0));
+    float2 c5 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 5 * Nx, 0));
+    float2 c6 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 6 * Nx, 0));
+
+    // Compute phi
+    float phi = (float)nx * exp_const;
+
+    // Multiply by twiddle factor
+    TWIDDLE_FACTOR_MULTIPLICATION(phi, c1);
+    TWIDDLE_FACTOR_MULTIPLICATION(2 * phi, c2);
+    TWIDDLE_FACTOR_MULTIPLICATION(3 * phi, c3);
+    TWIDDLE_FACTOR_MULTIPLICATION(4 * phi, c4);
+    TWIDDLE_FACTOR_MULTIPLICATION(5 * phi, c5);
+    TWIDDLE_FACTOR_MULTIPLICATION(6 * phi, c6);
+
+    // Compute DFT N = 7
+    DFT_7(c0, c1, c2, c3, c4, c5, c6);
+
+    // Store seven complex output values
+    vstore2(c0, 0, (__global float *)output.ptr);
+    vstore2(c1, 0, (__global float *)tensor3D_offset(&output, 0, Nx, 0));
+    vstore2(c2, 0, (__global float *)tensor3D_offset(&output, 0, 2 * Nx, 0));
+    vstore2(c3, 0, (__global float *)tensor3D_offset(&output, 0, 3 * Nx, 0));
+    vstore2(c4, 0, (__global float *)tensor3D_offset(&output, 0, 4 * Nx, 0));
+    vstore2(c5, 0, (__global float *)tensor3D_offset(&output, 0, 5 * Nx, 0));
+    vstore2(c6, 0, (__global float *)tensor3D_offset(&output, 0, 6 * Nx, 0));
+}
+
+/** Computes a stage of a radix-8 FFT on axis 0.
  *
  * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
@@ -937,14 +1606,14 @@ kernel void fft_radix_7_axis_0(
  * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
- * @param[out]    output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]     output_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]     output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]     output_stride_y                      Stride of the destination image in Y dimension (in bytes)
- * @param[in]     output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]     output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]     output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]     output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
  * @param[in]     Ni                                   Nx * Ny.
  * @param[in]     exp_const                            Exponent constant
@@ -1011,4 +1680,92 @@ kernel void fft_radix_8_axis_0(
     vstore2(c5, 0, (__global float *)tensor3D_offset(&output, 5 * Nx, 0, 0));
     vstore2(c6, 0, (__global float *)tensor3D_offset(&output, 6 * Nx, 0, 0));
     vstore2(c7, 0, (__global float *)tensor3D_offset(&output, 7 * Nx, 0, 0));
+}
+
+/** Computes a stage of a radix-8 FFT on axis 1.
+ *
+ * @note In order to perform the FFT function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @param[in,out] input_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in,out] input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in,out] input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in,out] input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in,out] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in,out] input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in,out] input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in,out] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out]    output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]     output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]     output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]     output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]     output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ * @param[in]     Nx                                   The butterfly span. Products of radix order of previous radix's stage
+ * @param[in]     Ni                                   Nx * Ny.
+ * @param[in]     exp_const                            Exponent constant
+ */
+kernel void fft_radix_8_axis_1(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+    ,
+    uint Nx, uint Ni, float exp_const)
+{
+    // Each work-item computes a single radix-8
+    uint kx = get_global_id(1);
+
+    // Compute nx
+    uint nx = kx % Nx;
+
+    // Compute n index
+    uint n = nx + (kx / Nx) * Ni;
+
+    // Get tensor pointers
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(input);
+    input.ptr += get_global_id(0) * input.stride_x + n * input.stride_y + get_global_id(2) * input.stride_z;
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
+    output.ptr += get_global_id(0) * output.stride_x + n * output.stride_y + get_global_id(2) * output.stride_z;
+#endif /* IN_PLACE */
+
+    // Load eight complex input values
+    float2 c0 = vload2(0, (__global float *)input.ptr);
+    float2 c1 = vload2(0, (__global float *)tensor3D_offset(&input, 0, Nx, 0));
+    float2 c2 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 2 * Nx, 0));
+    float2 c3 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 3 * Nx, 0));
+    float2 c4 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 4 * Nx, 0));
+    float2 c5 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 5 * Nx, 0));
+    float2 c6 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 6 * Nx, 0));
+    float2 c7 = vload2(0, (__global float *)tensor3D_offset(&input, 0, 7 * Nx, 0));
+
+    // Compute phi
+    float phi = (float)nx * exp_const;
+
+    // Multiply by twiddle factor
+    TWIDDLE_FACTOR_MULTIPLICATION(phi, c1);
+    TWIDDLE_FACTOR_MULTIPLICATION(2 * phi, c2);
+    TWIDDLE_FACTOR_MULTIPLICATION(3 * phi, c3);
+    TWIDDLE_FACTOR_MULTIPLICATION(4 * phi, c4);
+    TWIDDLE_FACTOR_MULTIPLICATION(5 * phi, c5);
+    TWIDDLE_FACTOR_MULTIPLICATION(6 * phi, c6);
+    TWIDDLE_FACTOR_MULTIPLICATION(7 * phi, c7);
+
+    // Compute DFT N = 8
+    DFT_8(c0, c1, c2, c3, c4, c5, c6, c7);
+
+    // Store eight complex output values
+    vstore2(c0, 0, (__global float *)output.ptr);
+    vstore2(c1, 0, (__global float *)tensor3D_offset(&output, 0, Nx, 0));
+    vstore2(c2, 0, (__global float *)tensor3D_offset(&output, 0, 2 * Nx, 0));
+    vstore2(c3, 0, (__global float *)tensor3D_offset(&output, 0, 3 * Nx, 0));
+    vstore2(c4, 0, (__global float *)tensor3D_offset(&output, 0, 4 * Nx, 0));
+    vstore2(c5, 0, (__global float *)tensor3D_offset(&output, 0, 5 * Nx, 0));
+    vstore2(c6, 0, (__global float *)tensor3D_offset(&output, 0, 6 * Nx, 0));
+    vstore2(c7, 0, (__global float *)tensor3D_offset(&output, 0, 7 * Nx, 0));
 }
 \ No newline at end of file
diff --git a/src/core/CL/cl_kernels/fft_digit_reverse.cl b/src/core/CL/cl_kernels/fft_digit_reverse.cl
new file mode 100644
index 0000000000..040c2846bd
--- /dev/null
+++ b/src/core/CL/cl_kernels/fft_digit_reverse.cl
@@ -0,0 +1,148 @@
+/*
+ * Copyright (c) 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(VEC_SIZE)
+/** Computes the digit reverse stage on axis X
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor 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 tensor 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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination tensor 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 tensor 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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in]  idx_ptr                           Pointer to the index tensor. Supported data types: U32
+ * @param[in]  idx_stride_x                      Stride of the index tensor in X dimension (in bytes)
+ * @param[in]  idx_step_x                        idx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  idx_offset_first_element_in_bytes The offset of the first element in the index tensor
+ */
+__kernel void fft_digit_reverse_axis_0(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    VECTOR_DECLARATION(idx))
+{
+    // Get tensor pointers
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+    Vector   idx = CONVERT_TO_VECTOR_STRUCT(idx);
+
+    const unsigned int iidx = *((__global uint *)(idx.ptr));
+
+    // Load data
+#if VEC_SIZE == 1
+    float data = *((__global float *)tensor3D_offset(&src, iidx, get_global_id(1), get_global_id(2)));
+#elif VEC_SIZE == 2
+    float2 data = vload2(0, (__global float *)tensor3D_offset(&src, iidx, get_global_id(1), get_global_id(2)));
+#else // VEC_SIZE == 1
+#error "vec_size of 1 and 2 are supported"
+#endif // VEC_SIZE == 1
+
+    // Create result
+#if VEC_SIZE == 1
+    float2 res = { data, 0 };
+#elif VEC_SIZE == 2
+    float2 res  = data;
+#else // VEC_SIZE == 1
+#error "vec_size of 1 and 2 are supported"
+#endif // VEC_SIZE == 1
+
+    // Store result
+#if defined(CONJ)
+    vstore2((float2)(res.s0, -res.s1), 0, (__global float *)dst.ptr);
+#else  // defined(CONJ)
+    vstore2(res, 0, (__global float *)dst.ptr);
+#endif // defined(CONJ)
+}
+
+/** Computes the digit reverse stage on axis Y
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor 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 tensor 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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination tensor 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 tensor 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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in]  idx_ptr                           Pointer to the index tensor. Supported data types: U32
+ * @param[in]  idx_stride_x                      Stride of the index tensor in X dimension (in bytes)
+ * @param[in]  idx_step_x                        idx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  idx_offset_first_element_in_bytes The offset of the first element in the index tensor
+ */
+__kernel void fft_digit_reverse_axis_1(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    VECTOR_DECLARATION(idx))
+{
+    // Get tensor pointers
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+    Vector   idx = CONVERT_TO_VECTOR_STRUCT_NO_STEP(idx);
+
+    const unsigned int iidx = *((__global uint *)vector_offset(&idx, (int)(get_global_id(1))));
+
+    // Load data
+#if VEC_SIZE == 1
+    float data = *((__global float *)tensor3D_offset(&src, get_global_id(0), iidx, get_global_id(2)));
+#elif VEC_SIZE == 2
+    float2 data = vload2(0, (__global float *)tensor3D_offset(&src, get_global_id(0), iidx, get_global_id(2)));
+#else // VEC_SIZE == 1
+#error "vec_size of 1 and 2 are supported"
+#endif // VEC_SIZE == 1
+
+    // Create result
+#if VEC_SIZE == 1
+    float2 res = { data, 0 };
+#elif VEC_SIZE == 2
+    float2 res  = data;
+#else // VEC_SIZE == 1
+#error "vec_size of 1 and 2 are supported"
+#endif // VEC_SIZE == 1
+
+    // Store result
+#if defined(CONJ)
+    vstore2((float2)(res.s0, -res.s1), 0, (__global float *)dst.ptr);
+#else  // defined(CONJ)
+    vstore2(res, 0, (__global float *)dst.ptr);
+#endif // defined(CONJ)
+}
+#endif // defined(VEC_SIZE)
+\ No newline at end of file
diff --git a/src/core/CL/cl_kernels/fft_scale.cl b/src/core/CL/cl_kernels/fft_scale.cl
new file mode 100644
index 0000000000..bf78a26eb8
--- /dev/null
+++ b/src/core/CL/cl_kernels/fft_scale.cl
@@ -0,0 +1,78 @@
+/*
+ * Copyright (c) 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"
+
+/** Computes the fft scale stage
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor 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 tensor 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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           (Optional) Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      (Optional) Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  dst_step_x                        (Optional) dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      (Optional) Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        (Optional) dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        (Optional) dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
+ * @param[in]  scale                             Scale to apply to the complex value
+ */
+__kernel void fft_scale_conj(
+    TENSOR3D_DECLARATION(src)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(dst)
+#endif /* not IN_PLACE */
+    ,
+    float scale)
+{
+    // Get tensor pointers
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+#if defined(IN_PLACE)
+    Tensor3D dst = src;
+#else  /* IN_PLACE */
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+#endif /* IN_PLACE */
+
+    // Store result
+#if VEC_SIZE == 1
+    *((__global float *)dst.ptr) = (*(__global float *)src.ptr) / scale;
+#elif VEC_SIZE == 2
+    // Load data
+    float2 data = vload2(0, (__global float *)src.ptr);
+    data /= scale;
+#if defined(CONJ)
+    vstore2((float2)(data.s0, -data.s1), 0, (__global float *)dst.ptr);
+#else  // defined(CONJ)
+    vstore2(data, 0, (__global float *)dst.ptr);
+#endif // defined(CONJ)
+#else  // VEC_SIZE == 1
+#error "vec_size of 1 and 2 are supported"
+#endif // VEC_SIZE == 1
+}
+\ No newline at end of file
diff --git a/src/core/CL/cl_kernels/pixelwise_mul_float.cl b/src/core/CL/cl_kernels/pixelwise_mul_float.cl
index 9fa540e946..d0e04b2ffe 100644
--- a/src/core/CL/cl_kernels/pixelwise_mul_float.cl
+++ b/src/core/CL/cl_kernels/pixelwise_mul_float.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -94,4 +94,52 @@ __kernel void pixelwise_mul_float(
     // Store result
     vstore16(res, 0, (__global DATA_TYPE_OUT *)out.ptr);
 }
-#endif /* defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT) */
-\ No newline at end of file
+#endif /* defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT) */
+
+/** Performs a pixelwise multiplication of complex float values
+ *
+ * @param[in]  in1_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in]  in1_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]  in1_step_x                        in1_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  in1_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in1_step_y                        in1_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in1_stride_z                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in1_step_z                        in1_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in1_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in]  in2_ptr                           Pointer to the source image. Supported data types: same as @p in1_ptr
+ * @param[in]  in2_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]  in2_step_x                        in2_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  in2_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in2_step_y                        in2_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in2_stride_z                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in2_step_z                        in2_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in2_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] out_ptr                           Pointer to the destination image. Supported data types: same as @p in1_ptr
+ * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
+ * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  out_stride_y                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_stride_z                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  out_step_z                        out_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+__kernel void pixelwise_mul_complex(
+    TENSOR3D_DECLARATION(in1),
+    TENSOR3D_DECLARATION(in2),
+    TENSOR3D_DECLARATION(out))
+{
+    // Get pixels pointer
+    Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
+    Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+
+    // Load data
+    float2 vin1 = vload2(0, (__global float *)in1.ptr);
+    float2 vin2 = vload2(0, (__global float *)in2.ptr);
+
+    // Perform complex multiplication
+    float2 res = { vin1.x *vin2.x - vin1.y * vin2.y, vin1.x *vin2.y + vin2.x * vin1.y };
+
+    // Store result
+    vstore2(res, 0, (__global float *)out.ptr);
+}
diff --git a/src/core/CL/cl_kernels/reduction_operation.cl b/src/core/CL/cl_kernels/reduction_operation.cl
index b4ede25296..2651123cf5 100644
--- a/src/core/CL/cl_kernels/reduction_operation.cl
+++ b/src/core/CL/cl_kernels/reduction_operation.cl
@@ -307,6 +307,10 @@ __kernel void reduction_operation_z(
     VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
     res = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
 
+#if defined(COMPLEX)
+    VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
+    res1 = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 8, 0, 0)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
+#endif // defined(COMPLEX)
 #if defined(SUM_SQUARE)
     res *= res;
 #endif // defined(SUM_SQUARE)
@@ -320,6 +324,11 @@ __kernel void reduction_operation_z(
         VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
         in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, z)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
 
+#if defined(COMPLEX)
+        VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
+        in1 = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 8, 0, z)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
+#endif // defined(COMPLEX)
+
 #if defined(ARG_MAX)
         uint16 cond_conv = CONVERT(isgreater(in, res), uint16);
         indx             = select(indx, z, cond_conv);
@@ -334,8 +343,11 @@ __kernel void reduction_operation_z(
 #endif // defined(SUM_SQUARE)
 #if defined(PROD)
         res *= in;
-#else  //!defined(PROD)
+#else //!defined(PROD)
         res += in;
+#if defined(COMPLEX)
+        res1 += in1;
+#endif // defined(COMPLEX)
 #endif //defined(PROD)
 #endif // defined(ARG_MAX) || defined(ARG_MIN)
     }
@@ -348,6 +360,9 @@ __kernel void reduction_operation_z(
     res /= DEPTH;
 #endif // defined(MEAN)
     vstore16(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)output.ptr);
+#if defined(COMPLEX)
+    vstore16(CONVERT(res1, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)tensor3D_offset(&output, 8, 0, 0));
+#endif // defined(COMPLEX)
 #endif // defined(ARG_MAX) || defined(ARG_MIN)
 }
 #endif /* defined(DEPTH) */
diff --git a/src/core/CL/kernels/CLFFTDigitReverseKernel.cpp b/src/core/CL/kernels/CLFFTDigitReverseKernel.cpp
index d72647c3c9..b04293db5b 100644
--- a/src/core/CL/kernels/CLFFTDigitReverseKernel.cpp
+++ b/src/core/CL/kernels/CLFFTDigitReverseKernel.cpp
@@ -34,16 +34,19 @@ namespace arm_compute
 {
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *idx, unsigned int axis)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *idx, const FFTDigitReverseKernelInfo &config)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 2, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() != DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_channels() != 1 && input->num_channels() != 2);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(idx, 1, DataType::U32);
-    ARM_COMPUTE_RETURN_ERROR_ON(axis != 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(std::set<unsigned int>({ 0, 1 }).count(config.axis) == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[config.axis] != idx->tensor_shape().x());
 
     // Checks performed when output is configured
     if((output != nullptr) && (output->total_size() != 0))
     {
+        ARM_COMPUTE_RETURN_ERROR_ON(output->num_channels() != 2);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }
@@ -51,11 +54,11 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *idx, unsigned int axis)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *idx, const FFTDigitReverseKernelInfo &config)
 {
-    ARM_COMPUTE_UNUSED(idx, axis);
+    ARM_COMPUTE_UNUSED(idx, config);
 
-    auto_init_if_empty(*output, *input);
+    auto_init_if_empty(*output, input->clone()->set_num_channels(2));
 
     Window win = calculate_max_window(*output, Steps());
     output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
@@ -69,25 +72,30 @@ CLFFTDigitReverseKernel::CLFFTDigitReverseKernel()
 {
 }
 
-void CLFFTDigitReverseKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *idx, unsigned int axis)
+void CLFFTDigitReverseKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *idx, const FFTDigitReverseKernelInfo &config)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, idx);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), idx->info(), axis));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), idx->info(), config));
 
     _input  = input;
     _output = output;
     _idx    = idx;
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("digit_reverse"));
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(input->info()->num_channels()));
+    build_opts.add_option_if(config.conjugate, "-DCONJ");
+    std::string kernel_name = "fft_digit_reverse_axis_" + support::cpp11::to_string(config.axis);
+    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info(), idx->info(), axis);
+    auto win_config = validate_and_configure_window(input->info(), output->info(), idx->info(), config);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure_internal(win_config.second);
 
     // Set config_id for enabling LWS tuning
-    _config_id = "digit_reverse_";
+    _config_id = kernel_name;
+    _config_id += "_";
     _config_id += lower_string(string_from_data_type(input->info()->data_type()));
     _config_id += "_";
     _config_id += support::cpp11::to_string(input->info()->dimension(0));
@@ -95,10 +103,10 @@ void CLFFTDigitReverseKernel::configure(const ICLTensor *input, ICLTensor *outpu
     _config_id += support::cpp11::to_string(input->info()->dimension(1));
 }
 
-Status CLFFTDigitReverseKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *idx, unsigned int axis)
+Status CLFFTDigitReverseKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *idx, const FFTDigitReverseKernelInfo &config)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, idx, axis));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), idx->clone().get(), axis).first);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, idx, config));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), idx->clone().get(), config).first);
 
     return Status{};
 }
diff --git a/src/core/CL/kernels/CLFFTRadixStageKernel.cpp b/src/core/CL/kernels/CLFFTRadixStageKernel.cpp
index 87a12b9da9..83d55b7092 100644
--- a/src/core/CL/kernels/CLFFTRadixStageKernel.cpp
+++ b/src/core/CL/kernels/CLFFTRadixStageKernel.cpp
@@ -38,12 +38,13 @@ namespace arm_compute
 {
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const FFTRadixStageKernelDescriptor &config)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const FFTRadixStageKernelInfo &config)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 2, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON(config.axis != 0);
     ARM_COMPUTE_RETURN_ERROR_ON(CLFFTRadixStageKernel::supported_radix().count(config.radix) == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(std::set<unsigned int>({ 0, 1 }).count(config.axis) == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[config.axis] % config.radix);
 
     // Checks performed when output is configured
     if((output != nullptr) && (output->total_size() != 0))
@@ -55,14 +56,18 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const FFTRadixStageKernelDescriptor &config)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const FFTRadixStageKernelInfo &config)
 {
     if(output != nullptr)
     {
         auto_init_if_empty(*output, *input);
     }
 
-    Window win = calculate_max_window(*input, Steps(config.radix));
+    // Setup window steps
+    Steps steps;
+    steps.set(config.axis, config.radix);
+
+    Window win = calculate_max_window(*input, steps);
     if(output != nullptr)
     {
         output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
@@ -77,7 +82,7 @@ CLFFTRadixStageKernel::CLFFTRadixStageKernel()
 {
 }
 
-void CLFFTRadixStageKernel::configure(ICLTensor *input, ICLTensor *output, const FFTRadixStageKernelDescriptor &config)
+void CLFFTRadixStageKernel::configure(ICLTensor *input, ICLTensor *output, const FFTRadixStageKernelInfo &config)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (output != nullptr) ? output->info() : nullptr, config));
@@ -105,7 +110,7 @@ void CLFFTRadixStageKernel::configure(ICLTensor *input, ICLTensor *output, const
         unsigned int       idx       = (1 + (_run_in_place ? 0 : 1)) * num_arguments_per_3D_tensor(); // Skip the input and output parameters
         _kernel.setArg<cl_uint>(idx++, config.Nx);
         _kernel.setArg<cl_uint>(idx++, Ni);
-        _kernel.setArg<cl_float>(idx++, exp_const);
+        _kernel.setArg<cl_float>(idx, exp_const);
     }
 
     // Configure kernel window
@@ -123,7 +128,7 @@ void CLFFTRadixStageKernel::configure(ICLTensor *input, ICLTensor *output, const
     _config_id += support::cpp11::to_string(input->info()->dimension(1));
 }
 
-Status CLFFTRadixStageKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const FFTRadixStageKernelDescriptor &config)
+Status CLFFTRadixStageKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const FFTRadixStageKernelInfo &config)
 {
     const bool run_in_place = (output == nullptr) || (output == input);
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, config));
diff --git a/src/core/CL/kernels/CLFFTScaleKernel.cpp b/src/core/CL/kernels/CLFFTScaleKernel.cpp
new file mode 100644
index 0000000000..59f1fd7502
--- /dev/null
+++ b/src/core/CL/kernels/CLFFTScaleKernel.cpp
@@ -0,0 +1,143 @@
+/*
+ * Copyright (c) 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/CLFFTScaleKernel.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/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Window.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 2, DataType::F32);
+
+    // Checks performed when output is configured
+    if((output != nullptr) && (output->total_size() != 0))
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(output->num_channels() != 1 && output->num_channels() != 2);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    // Configure kernel window
+    Window win = calculate_max_window(*input, Steps());
+
+    if(output != nullptr)
+    {
+        // Output auto inizialitation if not yet initialized
+        auto_init_if_empty(*output, *input->clone());
+
+        // CLFFTScaleKernel doesn't need padding so update_window_and_padding() can be skipped
+        Coordinates coord;
+        coord.set_num_dimensions(output->num_dimensions());
+        output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
+    }
+
+    return std::make_pair(Status{}, win);
+}
+} // namespace
+
+CLFFTScaleKernel::CLFFTScaleKernel()
+    : _input(nullptr), _output(nullptr), _run_in_place(false)
+{
+}
+
+void CLFFTScaleKernel::configure(ICLTensor *input, ICLTensor *output, const FFTScaleKernelInfo &config)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (output != nullptr) ? output->info() : nullptr));
+
+    _input        = input;
+    _output       = output;
+    _run_in_place = (output == nullptr) || (output == input);
+
+    // Create kernel
+    CLBuildOptions build_opts;
+    build_opts.add_option_if(_run_in_place, "-DIN_PLACE");
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(output != nullptr ? output->info()->num_channels() : input->info()->num_channels()));
+    build_opts.add_option_if(config.conjugate, "-DCONJ");
+    std::string kernel_name = "fft_scale_conj";
+    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+
+    // Set static arguments
+    unsigned int idx = (1 + (_run_in_place ? 0 : 1)) * num_arguments_per_3D_tensor(); // Skip the input and output parameters
+    _kernel.setArg<cl_float>(idx, config.scale);
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), _run_in_place ? nullptr : output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
+    // Set config_id for enabling LWS tuning
+    _config_id = kernel_name;
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_type(input->info()->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+}
+
+Status CLFFTScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const FFTScaleKernelInfo &config)
+{
+    ARM_COMPUTE_UNUSED(config);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
+
+    return Status{};
+}
+
+void CLFFTScaleKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+    Window slice     = collapsed.first_slice_window_3D();
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        if(!_run_in_place)
+        {
+            add_3D_tensor_argument(idx, _output, slice);
+        }
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(collapsed.slide_window_slice_3D(slice));
+}
+} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
index 286b94ebdc..9fa92bde75 100644
--- a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
+++ b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -38,8 +38,8 @@
 #include <set>
 #include <string>
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 namespace
 {
 constexpr unsigned int num_elems_processed_per_iteration = 16;
@@ -276,3 +276,139 @@ BorderSize CLPixelWiseMultiplicationKernel::border_size() const
     const unsigned int border        = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize);
     return BorderSize(0, border, 0, 0);
 }
+
+namespace
+{
+constexpr unsigned int num_elems_processed_per_iteration_complex = 1;
+
+Status validate_arguments_complex(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 2, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 2, DataType::F32);
+
+    const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape());
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
+
+    // Validate in case of configured output
+    if(output->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 2, DataType::F32);
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output");
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window_complex(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output)
+{
+    const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1, *input2);
+    const TensorShape &out_shape    = broadcast_pair.first;
+    const ValidRegion &valid_region = broadcast_pair.second;
+
+    // Auto initialize output if not initialized
+    const TensorInfo out_info(out_shape, input1->num_channels(), input1->data_type());
+    auto_init_if_empty(*output, out_info);
+
+    Window win        = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration_complex));
+    Window win_input1 = win.broadcast_if_dimension_le_one(*input1);
+    Window win_input2 = win.broadcast_if_dimension_le_one(*input2);
+
+    AccessWindowHorizontal input1_access(input1, 0, num_elems_processed_per_iteration_complex);
+    AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration_complex);
+    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_complex);
+
+    bool window_changed = update_window_and_padding(win_input1, input1_access)
+                          || update_window_and_padding(win_input2, input2_access)
+                          || update_window_and_padding(win, output_access);
+
+    output_access.set_valid_region(win, valid_region);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+CLComplexPixelWiseMultiplicationKernel::CLComplexPixelWiseMultiplicationKernel()
+    : _input1(nullptr), _input2(nullptr), _output(nullptr)
+{
+}
+
+void CLComplexPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_complex(input1->info(), input2->info(), output->info()));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window_complex(input1->info(), input2->info(), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+
+    _input1 = input1;
+    _input2 = input2;
+    _output = output;
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("pixelwise_mul_complex"));
+
+    ICLKernel::configure_internal(win_config.second);
+}
+
+Status CLComplexPixelWiseMultiplicationKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_complex(input1, input2, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_complex(input1->clone().get(), input2->clone().get(), output->clone().get()).first);
+
+    return Status{};
+}
+
+void CLComplexPixelWiseMultiplicationKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    const TensorShape &in_shape1 = _input1->info()->tensor_shape();
+    const TensorShape &in_shape2 = _input2->info()->tensor_shape();
+    const TensorShape &out_shape = _output->info()->tensor_shape();
+
+    bool can_collapse = true;
+    if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1)
+    {
+        can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ);
+        for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); ++d)
+        {
+            can_collapse = (in_shape1[d] == in_shape2[d]);
+        }
+    }
+
+    bool   has_collapsed = false;
+    Window collapsed     = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window;
+
+    const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1;
+    const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2;
+
+    Window slice        = collapsed.first_slice_window_3D();
+    Window slice_input1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed);
+    Window slice_input2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed);
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input1, slice_input1);
+        add_3D_tensor_argument(idx, _input2, slice_input2);
+        add_3D_tensor_argument(idx, _output, slice);
+        enqueue(queue, *this, slice);
+
+        collapsed.slide_window_slice_3D(slice_input1);
+        collapsed.slide_window_slice_3D(slice_input2);
+    }
+    while(collapsed.slide_window_slice_3D(slice));
+}
+
+BorderSize CLComplexPixelWiseMultiplicationKernel::border_size() const
+{
+    const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0));
+    const unsigned int border        = std::min<unsigned int>(num_elems_processed_per_iteration_complex - 1U, replicateSize);
+    return BorderSize(0, border, 0, 0);
+}
+} // namespace arm_compute
+\ No newline at end of file
diff --git a/src/core/CL/kernels/CLReductionOperationKernel.cpp b/src/core/CL/kernels/CLReductionOperationKernel.cpp
index 9f498b8273..db4850f14e 100644
--- a/src/core/CL/kernels/CLReductionOperationKernel.cpp
+++ b/src/core/CL/kernels/CLReductionOperationKernel.cpp
@@ -47,7 +47,14 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, u
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
+    if(input->num_channels() == 1)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 2, DataType::F32);
+    }
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(op == ReductionOperation::SUM_SQUARE && input->data_type() == DataType::QASYMM8, "Not supported reduction operation for QASYMM8");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
@@ -77,7 +84,7 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe
     output_shape.set(axis, 1);
     const bool is_arg_min_max   = (op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::ARG_IDX_MAX);
     DataType   output_data_type = is_arg_min_max ? DataType::U32 : input->data_type();
-    auto_init_if_empty(*output, output_shape, 1, output_data_type, input->quantization_info());
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape).set_data_type(output_data_type).reset_padding().set_is_resizable(true));
 
     const unsigned int num_elems_processed_per_iteration = (is_data_type_quantized(input->data_type()) && (axis == 0)) ? 1 : 16;
     Window             win                               = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
@@ -160,6 +167,7 @@ void CLReductionOperationKernel::configure(const ICLTensor *input, ICLTensor *ou
     build_opts.add_option_if(op == ReductionOperation::ARG_IDX_MAX, "-DARG_MAX");
     build_opts.add_option_if(op == ReductionOperation::ARG_IDX_MIN, "-DARG_MIN");
     build_opts.add_option_if(op == ReductionOperation::PROD, "-DPROD");
+    build_opts.add_option_if(input->info()->num_channels() == 2, "-DCOMPLEX");
 
     switch(op)
     {
diff --git a/src/runtime/CL/functions/CLFFT1D.cpp b/src/runtime/CL/functions/CLFFT1D.cpp
index d893cd3d1b..67111e7e5c 100644
--- a/src/runtime/CL/functions/CLFFT1D.cpp
+++ b/src/runtime/CL/functions/CLFFT1D.cpp
@@ -31,7 +31,7 @@
 namespace arm_compute
 {
 CLFFT1D::CLFFT1D(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _digit_reversed_input(), _digit_reverse_indices(), _digit_reverse_kernel(), _fft_kernels(), _num_ffts(0)
+    : _memory_group(std::move(memory_manager)), _digit_reverse_kernel(), _fft_kernels(), _scale_kernel(), _digit_reversed_input(), _digit_reverse_indices(), _num_ffts(0), _run_scale(false)
 {
 }
 
@@ -46,11 +46,18 @@ void CLFFT1D::configure(const ICLTensor *input, ICLTensor *output, const FFT1DIn
     const auto         decomposed_vector = arm_compute::helpers::fft::decompose_stages(N, supported_radix);
     ARM_COMPUTE_ERROR_ON(decomposed_vector.empty());
 
+    // Flags
+    _run_scale        = config.direction == FFTDirection::Inverse;
+    const bool is_c2r = input->info()->num_channels() == 2 && output->info()->num_channels() == 1;
+
     // Configure digit reverse
+    FFTDigitReverseKernelInfo digit_reverse_config;
+    digit_reverse_config.axis      = config.axis;
+    digit_reverse_config.conjugate = config.direction == FFTDirection::Inverse;
     TensorInfo digit_reverse_indices_info(TensorShape(input->info()->tensor_shape()[config.axis]), 1, DataType::U32);
     _digit_reverse_indices.allocator()->init(digit_reverse_indices_info);
     _memory_group.manage(&_digit_reversed_input);
-    _digit_reverse_kernel.configure(input, &_digit_reversed_input, &_digit_reverse_indices, config.axis);
+    _digit_reverse_kernel.configure(input, &_digit_reversed_input, &_digit_reverse_indices, digit_reverse_config);
 
     // Create and configure FFT kernels
     unsigned int Nx = 1;
@@ -60,16 +67,25 @@ void CLFFT1D::configure(const ICLTensor *input, ICLTensor *output, const FFT1DIn
     {
         const unsigned int radix_for_stage = decomposed_vector.at(i);
 
-        FFTRadixStageKernelDescriptor fft_kernel_desc;
-        fft_kernel_desc.axis           = config.axis;
-        fft_kernel_desc.radix          = radix_for_stage;
-        fft_kernel_desc.Nx             = Nx;
-        fft_kernel_desc.is_first_stage = (i == 0);
-        _fft_kernels[i].configure(&_digit_reversed_input, i == (_num_ffts - 1) ? output : nullptr, fft_kernel_desc);
+        FFTRadixStageKernelInfo fft_kernel_info;
+        fft_kernel_info.axis           = config.axis;
+        fft_kernel_info.radix          = radix_for_stage;
+        fft_kernel_info.Nx             = Nx;
+        fft_kernel_info.is_first_stage = (i == 0);
+        _fft_kernels[i].configure(&_digit_reversed_input, ((i == (_num_ffts - 1)) && !is_c2r) ? output : nullptr, fft_kernel_info);
 
         Nx *= radix_for_stage;
     }
 
+    // Configure scale kernel
+    if(_run_scale)
+    {
+        FFTScaleKernelInfo scale_config;
+        scale_config.scale     = static_cast<float>(N);
+        scale_config.conjugate = config.direction == FFTDirection::Inverse;
+        is_c2r ? _scale_kernel.configure(&_digit_reversed_input, output, scale_config) : _scale_kernel.configure(output, nullptr, scale_config);
+    }
+
     // Allocate tensors
     _digit_reversed_input.allocator()->allocate();
     _digit_reverse_indices.allocator()->allocate();
@@ -84,8 +100,9 @@ void CLFFT1D::configure(const ICLTensor *input, ICLTensor *output, const FFT1DIn
 Status CLFFT1D::validate(const ITensorInfo *input, const ITensorInfo *output, const FFT1DInfo &config)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 2, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON(config.axis != 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() != DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_channels() != 1 && input->num_channels() != 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(std::set<unsigned int>({ 0, 1 }).count(config.axis) == 0);
 
     // Check if FFT is decomposable
     const auto         supported_radix   = CLFFTRadixStageKernel::supported_radix();
@@ -96,6 +113,8 @@ Status CLFFT1D::validate(const ITensorInfo *input, const ITensorInfo *output, co
     // Checks performed when output is configured
     if((output != nullptr) && (output->total_size() != 0))
     {
+        ARM_COMPUTE_RETURN_ERROR_ON(output->num_channels() == 1 && input->num_channels() == 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(output->num_channels() != 1 && output->num_channels() != 2);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }
@@ -107,11 +126,19 @@ void CLFFT1D::run()
 {
     MemoryGroupResourceScope scope_mg(_memory_group);
 
+    // Run digit reverse
     CLScheduler::get().enqueue(_digit_reverse_kernel, false);
 
+    // Run radix kernels
     for(unsigned int i = 0; i < _num_ffts; ++i)
     {
-        CLScheduler::get().enqueue(_fft_kernels[i], i == (_num_ffts - 1));
+        CLScheduler::get().enqueue(_fft_kernels[i], i == (_num_ffts - 1) && !_run_scale);
+    }
+
+    // Run output scaling
+    if(_run_scale)
+    {
+        CLScheduler::get().enqueue(_scale_kernel, true);
     }
 }
 } // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLFFT2D.cpp b/src/runtime/CL/functions/CLFFT2D.cpp
new file mode 100644
index 0000000000..4300fb4e32
--- /dev/null
+++ b/src/runtime/CL/functions/CLFFT2D.cpp
@@ -0,0 +1,95 @@
+/*
+ * Copyright (c) 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/CLFFT2D.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+
+namespace arm_compute
+{
+CLFFT2D::CLFFT2D(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(memory_manager), _first_pass_func(memory_manager), _second_pass_func(memory_manager), _first_pass_tensor()
+{
+}
+
+void CLFFT2D::configure(const ICLTensor *input, ICLTensor *output, const FFT2DInfo &config)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(CLFFT2D::validate(input->info(), output->info(), config));
+
+    // Setup first pass
+    FFT1DInfo first_pass_config;
+    first_pass_config.axis      = config.axes.first;
+    first_pass_config.direction = config.direction;
+    _memory_group.manage(&_first_pass_tensor);
+    _first_pass_func.configure(input, &_first_pass_tensor, first_pass_config);
+
+    // Setup second pass
+    FFT1DInfo second_pass_config;
+    second_pass_config.axis      = config.axes.second;
+    second_pass_config.direction = config.direction;
+    _second_pass_func.configure(&_first_pass_tensor, output, second_pass_config);
+    _first_pass_tensor.allocator()->allocate();
+}
+
+Status CLFFT2D::validate(const ITensorInfo *input, const ITensorInfo *output, const FFT2DInfo &config)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+
+    // Create intermediate tensor info
+    TensorInfo first_pass_tensor(input->clone()->set_is_resizable(true).reset_padding().set_num_channels(2));
+
+    // Validate first pass
+    FFT1DInfo first_pass_config;
+    first_pass_config.axis      = config.axes.first;
+    first_pass_config.direction = config.direction;
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFFT1D::validate(input, &first_pass_tensor, first_pass_config));
+
+    // Validate second pass
+    FFT1DInfo second_pass_config;
+    second_pass_config.axis      = config.axes.second;
+    second_pass_config.direction = config.direction;
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFFT1D::validate(&first_pass_tensor, output, second_pass_config));
+
+    // Checks performed when output is configured
+    if((output != nullptr) && (output->total_size() != 0))
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+void CLFFT2D::run()
+{
+    _memory_group.acquire();
+
+    _first_pass_func.run();
+    _second_pass_func.run();
+
+    _memory_group.release();
+}
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLFFTConvolutionLayer.cpp b/src/runtime/CL/functions/CLFFTConvolutionLayer.cpp
new file mode 100644
index 0000000000..441c1c7214
--- /dev/null
+++ b/src/runtime/CL/functions/CLFFTConvolutionLayer.cpp
@@ -0,0 +1,380 @@
+/*
+ * Copyright (c) 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/CLFFTConvolutionLayer.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/helpers/fft.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+#include "arm_compute/runtime/CPP/CPPScheduler.h"
+
+namespace arm_compute
+{
+namespace
+{
+int pad_decomposable(int N)
+{
+    const auto supported_radix = CLFFTRadixStageKernel::supported_radix();
+
+    int  pad           = 0;
+    bool is_decomposed = false;
+    while(!is_decomposed)
+    {
+        const auto decomposed_vector = arm_compute::helpers::fft::decompose_stages(N++, supported_radix);
+        is_decomposed                = !decomposed_vector.empty();
+        if(!is_decomposed)
+        {
+            ++pad;
+        }
+    }
+    return pad;
+}
+} // namespace
+CLFFTConvolutionLayer::CLFFTConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(memory_manager),
+      _flip_weights_func(),
+      _permute_input_func(),
+      _permute_output_func(),
+      _permute_weights_func(),
+      _permute_bias_func(),
+      _pad_input_func(),
+      _pad_weights_func(),
+      _transform_input_func(memory_manager),
+      _transform_weights_func(memory_manager),
+      _itransform_output_func(memory_manager),
+      _prod_func(),
+      _reduce_func(),
+      _extract_output_func(),
+      _bias_add_func(),
+      _activation_layer_func(),
+      _permuted_input(),
+      _permuted_weights(),
+      _permuted_bias(),
+      _permuted_output(),
+      _padded_input(),
+      _padded_weights(),
+      _flip_axis(),
+      _flipped_weights(),
+      _transformed_input(),
+      _transformed_weights(),
+      _input_weights_product(),
+      _output_product(),
+      _output_reduced(),
+      _itransformed_output(),
+      _reshaped_output(),
+      _bias_output(),
+      _original_weights(nullptr),
+      _original_bias(nullptr),
+      _is_activationlayer_enabled(false),
+      _needs_permute(false),
+      _has_bias(false),
+      _is_prepared(false)
+{
+}
+
+void CLFFTConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
+                                      const ActivationLayerInfo &act_info)
+{
+    _original_weights = weights;
+    _original_bias    = biases;
+
+    // Flat if bias addition is required
+    _has_bias = biases != nullptr;
+
+    // Get indices for the width and height
+    const size_t idx_width  = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
+
+    // Input shape, kernel size and output tile
+    const Size2D input_dims  = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]);
+    const Size2D kernel_size = Size2D(weights->info()->tensor_shape()[idx_width], weights->info()->tensor_shape()[idx_height]);
+    const Size2D pad_valid   = Size2D(pad_decomposable(input_dims.x() + kernel_size.x() - 1),
+                                      pad_decomposable(input_dims.y() + kernel_size.y() - 1));
+    // Tensors to use
+    ICLTensor       *input_to_use   = input;
+    const ICLTensor *weights_to_use = weights;
+    ICLTensor       *output_to_use  = _has_bias ? &_bias_output : output;
+
+    // Permute bias
+    _permute_bias_func.configure(biases, &_permuted_bias, PermutationVector(1U, 2U, 0U));
+    _permuted_bias.info()->set_data_layout(DataLayout::NCHW);
+
+    // Permute input if needed
+    _needs_permute = input->info()->data_layout() == DataLayout::NHWC;
+    if(_needs_permute)
+    {
+        _memory_group.manage(&_permuted_input);
+        // Configure the function to transform the input tensor from NHWC -> NCHW
+        _permute_input_func.configure(input, &_permuted_input, PermutationVector(1U, 2U, 0U));
+        _permuted_input.info()->set_data_layout(DataLayout::NCHW);
+
+        // Configure the function to transform the weights tensor from HWI -> IHW
+        _permute_weights_func.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
+        _permuted_weights.info()->set_data_layout(DataLayout::NCHW);
+
+        input_to_use   = &_permuted_input;
+        weights_to_use = &_permuted_weights;
+    }
+
+    // Flip weights
+    _flipped_weights.allocator()->init(weights_to_use->info()->clone()->set_is_resizable(true).reset_padding());
+    _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32));
+    _flip_weights_func.configure(weights_to_use, &_flipped_weights, &_flip_axis);
+
+    // Pad weights
+    const PaddingList padding_w = { { 0, input_dims.x() + pad_valid.x() - 1 }, { 0, input_dims.y() + pad_valid.y() - 1 } };
+    _pad_weights_func.configure(&_flipped_weights, &_padded_weights, padding_w);
+
+    // Transform weights
+    _transform_weights_func.configure(&_padded_weights, &_transformed_weights, FFT2DInfo());
+
+    // Pad input
+    const PaddingList padding_in = { { 0, kernel_size.x() + pad_valid.x() - 1 }, { 0, kernel_size.y() + pad_valid.y() - 1 } };
+    _memory_group.manage(&_padded_input);
+    _pad_input_func.configure(input_to_use, &_padded_input, padding_in);
+    if(_needs_permute)
+    {
+        _permuted_input.allocator()->allocate();
+    }
+
+    // Transform input
+    _memory_group.manage(&_transformed_input);
+    _transform_input_func.configure(&_padded_input, &_transformed_input, FFT2DInfo());
+    _padded_input.allocator()->allocate();
+
+    // Perform product
+    _memory_group.manage(&_output_product);
+    _prod_func.configure(&_transformed_input, &_transformed_weights, &_output_product);
+    _transformed_input.allocator()->allocate();
+
+    // Perform reduction
+    _memory_group.manage(&_output_reduced);
+    _reduce_func.configure(&_output_product, &_output_reduced, 2, ReductionOperation::SUM);
+    _output_product.allocator()->allocate();
+
+    // Transform output
+    _memory_group.manage(&_itransformed_output);
+    FFT2DInfo itranform_info;
+    itranform_info.direction = FFTDirection::Inverse;
+    _itransformed_output.allocator()->init(_output_reduced.info()->clone()->set_is_resizable(true).set_num_channels(1).reset_padding());
+    _itransform_output_func.configure(&_output_reduced, &_itransformed_output, itranform_info);
+    _output_reduced.allocator()->allocate();
+
+    // Reshape output
+    TensorShape reshaped_shape = _itransformed_output.info()->tensor_shape();
+    reshaped_shape.remove_dimension(2);
+    _reshaped_output.allocator()->init(_itransformed_output.info()->clone()->set_tensor_shape(reshaped_shape));
+
+    // Extract correct region
+    const int start_left = kernel_size.x() - conv_info.pad_left() - 1;
+    const int start_top  = kernel_size.y() - conv_info.pad_top() - 1;
+    const int end_right  = _reshaped_output.info()->tensor_shape().x() - (kernel_size.x() - conv_info.pad_right() - 1) - pad_valid.x();
+    const int end_botton = _reshaped_output.info()->tensor_shape().y() - (kernel_size.y() - conv_info.pad_bottom() - 1) - pad_valid.y();
+    if(_has_bias)
+    {
+        _memory_group.manage(&_bias_output);
+    }
+    else if(_needs_permute)
+    {
+        output_to_use = &_permuted_output;
+        _memory_group.manage(&_permuted_output);
+    }
+    _extract_output_func.configure(&_reshaped_output, output_to_use, Coordinates(start_left, start_top), Coordinates(end_right, end_botton));
+    _itransformed_output.allocator()->allocate();
+
+    // Add bias
+    if(biases != nullptr)
+    {
+        output_to_use = output;
+        if(_needs_permute)
+        {
+            output_to_use = &_permuted_output;
+            _memory_group.manage(&_permuted_output);
+        }
+        auto_init_if_empty(*output_to_use->info(), *_bias_output.info());
+        _bias_add_func.configure(&_bias_output, &_permuted_bias, output_to_use, ConvertPolicy::WRAP);
+        _bias_output.allocator()->allocate();
+    }
+
+    // Permute output
+    if(_needs_permute)
+    {
+        // Configure the function to transform the convoluted output to ACL's native ordering format NCHW
+        _permuted_output.info()->set_data_layout(DataLayout::NCHW);
+        _permute_output_func.configure(&_permuted_output, output, PermutationVector(2U, 0U, 1U));
+
+        // Allocate tensors
+        _permuted_output.allocator()->allocate();
+    }
+
+    // Configure Activation Layer
+    _is_activationlayer_enabled = act_info.enabled();
+    if(_is_activationlayer_enabled)
+    {
+        _activation_layer_func.configure(output, nullptr, act_info);
+    }
+
+    // Setup flip axis data
+    _flip_axis.allocator()->allocate();
+    _flip_axis.map(true);
+    auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer());
+    axis_data[0]   = 0;
+    axis_data[1]   = 1;
+    _flip_axis.unmap();
+}
+
+Status CLFFTConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                       const ActivationLayerInfo &act_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+
+    // Get indices for the width and height
+    const size_t idx_width  = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
+
+    // Input shape, kernel size and output tile
+    const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]);
+
+    // Strides
+    const auto strides = conv_info.stride();
+    ARM_COMPUTE_RETURN_ERROR_ON(strides.first != strides.second && strides.first != 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(kernel_size.x() != kernel_size.y());
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_left() != (kernel_size.x() / 2) || conv_info.pad_right() != (kernel_size.x() / 2));
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_top() != (kernel_size.y() / 2) || conv_info.pad_bottom() != (kernel_size.y() / 2));
+
+    // Validate biases
+    if(biases != nullptr)
+    {
+        const size_t idx_channels = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_channels] != biases->tensor_shape().x());
+    }
+
+    // Checks performed when output is configured
+    if((output != nullptr) && (output->total_size() != 0))
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+
+        // Validate Activation Layer
+        if(act_info.enabled())
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output, nullptr, act_info));
+        }
+    }
+
+    return Status{};
+}
+
+void CLFFTConvolutionLayer::run()
+{
+    prepare();
+
+    _memory_group.acquire();
+
+    // Transform input
+    if(_needs_permute)
+    {
+        _permute_input_func.run();
+    }
+    _pad_input_func.run();
+    _transform_input_func.run();
+
+    // Perform operations to frequency domain
+    _prod_func.run();
+    _reduce_func.run();
+
+    // Transform output
+    _itransform_output_func.run();
+    _reshaped_output.allocator()->import_memory(_itransformed_output.cl_buffer());
+    _extract_output_func.run();
+    // Add bias
+    if(_has_bias)
+    {
+        _bias_add_func.run();
+    }
+    if(_needs_permute)
+    {
+        _permute_output_func.run();
+    }
+
+    // Run activation layer
+    if(_is_activationlayer_enabled)
+    {
+        _activation_layer_func.run();
+    }
+
+    _memory_group.release();
+}
+
+void CLFFTConvolutionLayer::prepare()
+{
+    if(!_is_prepared)
+    {
+        // Permute bias to NCHW
+        if(_original_bias != nullptr)
+        {
+            _permuted_bias.allocator()->allocate();
+            _permute_bias_func.run();
+            _original_bias->mark_as_unused();
+        }
+
+        const ICLTensor *cur_weights = _original_weights;
+        // Permute weights
+        if(_needs_permute)
+        {
+            ARM_COMPUTE_ERROR_ON(!cur_weights->is_used());
+
+            _permuted_weights.allocator()->allocate();
+            _permute_weights_func.run();
+            cur_weights->mark_as_unused();
+            cur_weights = &_permuted_weights;
+        }
+
+        // Flip weights
+        _flipped_weights.allocator()->allocate();
+        _flip_weights_func.run();
+        cur_weights->mark_as_unused();
+
+        // Pad weights
+        _padded_weights.allocator()->allocate();
+        _pad_weights_func.run();
+        _flipped_weights.mark_as_unused();
+        CLScheduler::get().queue().finish();
+        _flipped_weights.allocator()->free();
+
+        // Transform weights to frequence domain
+        _transformed_weights.allocator()->allocate();
+        _transform_weights_func.run();
+        _padded_weights.mark_as_unused();
+        CLScheduler::get().queue().finish();
+        _padded_weights.allocator()->free();
+
+        _is_prepared = true;
+    }
+}
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp b/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp
index b4c20db3da..959464ce14 100644
--- a/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp
+++ b/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,8 +29,8 @@
 
 #include <utility>
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 void CLPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, float scale,
                                           ConvertPolicy overflow_policy, RoundingPolicy rounding_policy)
 {
@@ -54,3 +54,26 @@ Status CLPixelWiseMultiplication::validate(const ITensorInfo *input1, const ITen
 {
     return CLPixelWiseMultiplicationKernel::validate(input1, input2, output, scale, overflow_policy, rounding_policy);
 }
+
+void CLComplexPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLComplexPixelWiseMultiplicationKernel>();
+    k->configure(input1, input2, output);
+    _kernel = std::move(k);
+
+    if(output->info()->dimension(0) > 1)
+    {
+        ICLTensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2;
+
+        if(broadcasted_info->info()->dimension(0) == 1)
+        {
+            _border_handler.configure(broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE);
+        }
+    }
+}
+
+Status CLComplexPixelWiseMultiplication::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
+{
+    return CLComplexPixelWiseMultiplicationKernel::validate(input1, input2, output);
+}
+} // namespace arm_compute
+\ No newline at end of file
diff --git a/tests/benchmark/CL/ConvolutionLayer.cpp b/tests/benchmark/CL/ConvolutionLayer.cpp
index 5eb33658ff..20828b7717 100644
--- a/tests/benchmark/CL/ConvolutionLayer.cpp
+++ b/tests/benchmark/CL/ConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,6 +29,7 @@
 #include "arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h"
 #include "tests/CL/CLAccessor.h"
 #include "tests/benchmark/fixtures/ConvolutionLayerFixture.h"
+#include "tests/benchmark/fixtures/FFTConvolutionLayerFixture.h"
 #include "tests/benchmark/fixtures/WinogradConvolutionLayerFixture.h"
 #include "tests/datasets/system_tests/alexnet/AlexNetConvolutionLayerDataset.h"
 #include "tests/datasets/system_tests/googlenet/inceptionv1/GoogLeNetInceptionV1ConvolutionLayerDataset.h"
@@ -41,6 +42,9 @@
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "utils/TypePrinter.h"
+#include <arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h>
+#include <tests/datasets/SmallConvolutionLayerDataset.h>
+#include <tests/datasets/system_tests/resnet12/ResNet12ConvolutionLayerDataset.h>
 
 namespace arm_compute
 {
@@ -53,11 +57,17 @@ namespace
 const auto data_types = framework::dataset::make("DataType", { DataType::F16, DataType::F32, DataType::QASYMM8 });
 } // namespace
 
-using CLGEMMConvolutionLayerFixture = ConvolutionLayerFixture<CLTensor, CLGEMMConvolutionLayer, CLAccessor>;
-
 TEST_SUITE(CL)
 
-using CLWinogradLayerFixture = WinogradConvolutionLayerFixture<CLTensor, CLWinogradConvolutionLayer, CLAccessor>;
+using CLGEMMConvolutionLayerFixture = ConvolutionLayerFixture<CLTensor, CLGEMMConvolutionLayer, CLAccessor>;
+using CLWinogradLayerFixture        = WinogradConvolutionLayerFixture<CLTensor, CLWinogradConvolutionLayer, CLAccessor>;
+using CLFFTConvolutionLayerFixture  = FFTConvolutionLayerFixture<CLTensor, CLFFTConvolutionLayer, CLAccessor>;
+
+REGISTER_FIXTURE_DATA_TEST_CASE(ResNet12FFTLayer, CLFFTConvolutionLayerFixture, framework::DatasetMode::ALL,
+                                framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::ResNet12FFTConvolutionLayerDataset(),
+                                                                                                                    framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))),
+                                                                                        framework::dataset::make("DataType", { DataType::F32 })),
+                                                            framework::dataset::make("Batches", 1)));
 
 REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetWinogradLayer, CLWinogradLayerFixture, framework::DatasetMode::ALL,
                                 framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::AlexNetWinogradLayerDataset(),
diff --git a/tests/benchmark/CL/FFT.cpp b/tests/benchmark/CL/FFT.cpp
index b345d58eaf..7f1ae63708 100644
--- a/tests/benchmark/CL/FFT.cpp
+++ b/tests/benchmark/CL/FFT.cpp
@@ -24,6 +24,7 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/CL/functions/CLFFT1D.h"
+#include "arm_compute/runtime/CL/functions/CLFFT2D.h"
 #include "tests/CL/CLAccessor.h"
 #include "tests/benchmark/fixtures/FFTFixture.h"
 #include "tests/framework/Macros.h"
@@ -42,13 +43,17 @@ const auto data_types = framework::dataset::make("DataType", { DataType::F32 });
 const auto shapes     = framework::dataset::make("Shapes", { TensorShape(192U, 128U, 64U), TensorShape(224U, 224U) });
 } // namespace
 
-using CLFFT1DFixture = FFT1DFixture<CLTensor, CLFFT1D, CLAccessor>;
+using CLFFT1DFixture = FFTFixture<CLTensor, CLFFT1D, FFT1DInfo, CLAccessor>;
+using CLFFT2DFixture = FFTFixture<CLTensor, CLFFT2D, FFT2DInfo, CLAccessor>;
 
 TEST_SUITE(CL)
 
 REGISTER_FIXTURE_DATA_TEST_CASE(FFT1D, CLFFT1DFixture, framework::DatasetMode::ALL,
                                 framework::dataset::combine(shapes, data_types));
 
+REGISTER_FIXTURE_DATA_TEST_CASE(FFT2D, CLFFT2DFixture, framework::DatasetMode::ALL,
+                                framework::dataset::combine(shapes, data_types));
+
 TEST_SUITE_END() // CL
 } // namespace benchmark
 } // namespace test
diff --git a/tests/benchmark/fixtures/FFTConvolutionLayerFixture.h b/tests/benchmark/fixtures/FFTConvolutionLayerFixture.h
new file mode 100644
index 0000000000..2c53e3ad9b
--- /dev/null
+++ b/tests/benchmark/fixtures/FFTConvolutionLayerFixture.h
@@ -0,0 +1,100 @@
+/*
+ * Copyright (c) 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_FFT_CONVOLUTION_LAYER_FIXTURE
+#define ARM_COMPUTE_TEST_FFT_CONVOLUTION_LAYER_FIXTURE
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "tests/Globals.h"
+#include "tests/Utils.h"
+#include "tests/framework/Fixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace benchmark
+{
+/** Fixture that can be used for NEON and CL */
+template <typename TensorType, typename Function, typename Accessor>
+class FFTConvolutionLayerFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape src_shape, TensorShape weights_shape, TensorShape biases_shape, TensorShape dst_shape, PadStrideInfo info, Size2D dilation, ActivationLayerInfo act_info, DataType data_type,
+               int batches)
+    {
+        ARM_COMPUTE_UNUSED(dilation);
+
+        // Set batched in source and destination shapes
+
+        src_shape.set(3 /* batch */, batches);
+        dst_shape.set(3 /* batch */, batches);
+
+        // Create tensors
+        src     = create_tensor<TensorType>(src_shape, data_type, 1);
+        weights = create_tensor<TensorType>(weights_shape, data_type, 1);
+        biases  = create_tensor<TensorType>(biases_shape, data_type, 1);
+        dst     = create_tensor<TensorType>(dst_shape, data_type, 1);
+
+        // Create and configure function
+        conv_layer.configure(&src, &weights, &biases, &dst, info, act_info);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        weights.allocator()->allocate();
+        biases.allocator()->allocate();
+        dst.allocator()->allocate();
+    }
+
+    void run()
+    {
+        conv_layer.run();
+    }
+
+    void sync()
+    {
+        sync_if_necessary<TensorType>();
+        sync_tensor_if_necessary<TensorType>(dst);
+    }
+
+    void teardown()
+    {
+        src.allocator()->free();
+        weights.allocator()->free();
+        biases.allocator()->free();
+        dst.allocator()->free();
+    }
+
+private:
+    TensorType src{};
+    TensorType weights{};
+    TensorType biases{};
+    TensorType dst{};
+    Function   conv_layer{};
+};
+} // namespace benchmark
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_FFT_CONVOLUTION_LAYER_FIXTURE */
diff --git a/tests/benchmark/fixtures/FFTFixture.h b/tests/benchmark/fixtures/FFTFixture.h
index c9c4e3a88e..53897b1b14 100644
--- a/tests/benchmark/fixtures/FFTFixture.h
+++ b/tests/benchmark/fixtures/FFTFixture.h
@@ -36,8 +36,8 @@ namespace test
 {
 namespace benchmark
 {
-template <typename TensorType, typename Function, typename Accessor>
-class FFT1DFixture : public framework::Fixture
+template <typename TensorType, typename Function, typename FFTInfo, typename Accessor>
+class FFTFixture : public framework::Fixture
 {
 public:
     template <typename...>
@@ -48,7 +48,7 @@ public:
         dst = create_tensor<TensorType>(shape, data_type, 2);
 
         // Create and configure function
-        fft_func.configure(&src, &dst, FFT1DInfo());
+        fft_func.configure(&src, &dst, FFTInfo());
 
         // Allocate tensors
         src.allocator()->allocate();
diff --git a/tests/datasets/SmallConvolutionLayerDataset.h b/tests/datasets/SmallConvolutionLayerDataset.h
index 73f1554c49..22d0bc582a 100644
--- a/tests/datasets/SmallConvolutionLayerDataset.h
+++ b/tests/datasets/SmallConvolutionLayerDataset.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -132,6 +132,17 @@ public:
     }
 };
 
+class SmallFFTConvolutionLayerDataset final : public ConvolutionLayerDataset
+{
+public:
+    SmallFFTConvolutionLayerDataset()
+    {
+        add_config(TensorShape(8U, 7U, 3U), TensorShape(3U, 3U, 3U, 2U), TensorShape(2U), TensorShape(8U, 7U, 2U), PadStrideInfo(1, 1, 1, 1));
+        add_config(TensorShape(64U, 32U, 5U), TensorShape(5U, 5U, 5U, 10U), TensorShape(10U), TensorShape(64U, 32U, 10U), PadStrideInfo(1, 1, 2, 2));
+        add_config(TensorShape(192U, 128U, 8U), TensorShape(9U, 9U, 8U, 3U), TensorShape(3U), TensorShape(192U, 128U, 3U), PadStrideInfo(1, 1, 4, 4));
+    }
+};
+
 class SmallConvolutionLayerDataset final : public ConvolutionLayerDataset
 {
 public:
diff --git a/tests/datasets/system_tests/resnet12/ResNet12ConvolutionLayerDataset.h b/tests/datasets/system_tests/resnet12/ResNet12ConvolutionLayerDataset.h
new file mode 100644
index 0000000000..b960dceafd
--- /dev/null
+++ b/tests/datasets/system_tests/resnet12/ResNet12ConvolutionLayerDataset.h
@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 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_RESNET12_CONVOLUTION_LAYER_DATASET
+#define ARM_COMPUTE_TEST_RESNET12_CONVOLUTION_LAYER_DATASET
+
+#include "tests/datasets/ConvolutionLayerDataset.h"
+
+#include "utils/TypePrinter.h"
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace datasets
+{
+class ResNet12FFTConvolutionLayerDataset final : public ConvolutionLayerDataset
+{
+public:
+    ResNet12FFTConvolutionLayerDataset()
+    {
+        add_config(TensorShape(192U, 128U, 64U), TensorShape(9U, 9U, 64U, 3U), TensorShape(3U), TensorShape(192U, 128U, 3U), PadStrideInfo(1, 1, 4, 4));
+    }
+};
+} // namespace datasets
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_RESNET12_CONVOLUTION_LAYER_DATASET */
diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp
index 41d2b7bb5e..f1f9b59330 100644
--- a/tests/validation/CL/ConvolutionLayer.cpp
+++ b/tests/validation/CL/ConvolutionLayer.cpp
@@ -46,7 +46,7 @@ namespace validation
 namespace
 {
 constexpr AbsoluteTolerance<float>  absolute_tolerance_float(0.0001f);    /**< Absolute Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
-RelativeTolerance<float>            tolerance_f32(0.05f);                 /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+RelativeTolerance<float>            tolerance_f32(0.1f);                  /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
 RelativeTolerance<half_float::half> tolerance_f16(half_float::half(0.2)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
 constexpr AbsoluteTolerance<float>  tolerance_qasymm8(1);                 /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
 constexpr float                     tolerance_num = 0.07f;                /**< Tolerance number */
diff --git a/tests/validation/CL/FFT.cpp b/tests/validation/CL/FFT.cpp
index 0d29532c29..9fdd85b604 100644
--- a/tests/validation/CL/FFT.cpp
+++ b/tests/validation/CL/FFT.cpp
@@ -24,7 +24,10 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/CL/functions/CLFFT1D.h"
+#include "arm_compute/runtime/CL/functions/CLFFT2D.h"
+#include "arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h"
 #include "tests/CL/CLAccessor.h"
+#include "tests/datasets/SmallConvolutionLayerDataset.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
@@ -40,7 +43,7 @@ namespace validation
 namespace
 {
 const auto data_types = framework::dataset::make("DataType", { DataType::F32 });
-const auto shapes     = framework::dataset::make("TensorShape", { TensorShape(2U, 2U, 3U), TensorShape(3U, 2U, 3U),
+const auto shapes_1d  = framework::dataset::make("TensorShape", { TensorShape(2U, 2U, 3U), TensorShape(3U, 2U, 3U),
                                                                   TensorShape(4U, 2U, 3U), TensorShape(5U, 2U, 3U),
                                                                   TensorShape(7U, 2U, 3U), TensorShape(8U, 2U, 3U),
                                                                   TensorShape(9U, 2U, 3U), TensorShape(25U, 2U, 3U),
@@ -48,11 +51,27 @@ const auto shapes     = framework::dataset::make("TensorShape", { TensorShape(2U
                                                                   TensorShape(16U, 2U, 3U), TensorShape(32U, 2U, 3U),
                                                                   TensorShape(96U, 2U, 2U)
                                                                 });
+const auto shapes_2d = framework::dataset::make("TensorShape", { TensorShape(2U, 2U, 3U), TensorShape(3U, 6U, 3U),
+                                                                 TensorShape(4U, 5U, 3U), TensorShape(5U, 7U, 3U),
+                                                                 TensorShape(7U, 25U, 3U), TensorShape(8U, 2U, 3U),
+                                                                 TensorShape(9U, 16U, 3U), TensorShape(25U, 32U, 3U),
+                                                                 TensorShape(192U, 128U, 2U)
+                                                               });
+
+const auto ActivationFunctionsSmallDataset = framework::dataset::make("ActivationInfo",
+{
+    ActivationLayerInfo(),
+    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f)
+});
+
+RelativeTolerance<float> tolerance_f32(0.1f);   /**< Relative tolerance value for FP32 */
+constexpr float          tolerance_num = 0.07f; /**< Tolerance number */
+
 } // namespace
 TEST_SUITE(CL)
 TEST_SUITE(FFT1D)
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(shapes, data_types),
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(shapes_1d, data_types),
                shape, data_type)
 {
     // Create tensors
@@ -81,19 +100,19 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(shapes, data_
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
         framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32), // Mismatching data types
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32), // Mismatching shapes
-                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Invalid channels
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 3, DataType::F32), // Invalid channels
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32), // Unsupported axis
                                                 TensorInfo(TensorShape(11U, 13U, 2U), 2, DataType::F32), // Undecomposable FFT
                                                 TensorInfo(TensorShape(25U, 13U, 2U), 2, DataType::F32),
         }),
         framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F16),
                                                 TensorInfo(TensorShape(16U, 13U, 2U), 2, DataType::F32),
-                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32),
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32),
                                                 TensorInfo(TensorShape(11U, 13U, 2U), 2, DataType::F32),
                                                 TensorInfo(TensorShape(25U, 13U, 2U), 2, DataType::F32),
         })),
-        framework::dataset::make("Axis", { 0, 0, 0, 1, 0, 0 })),
+        framework::dataset::make("Axis", { 0, 0, 0, 2, 0, 0 })),
         framework::dataset::make("Expected", { false, false, false, false, false, true })),
         input_info, output_info, axis, expected)
 {
@@ -106,19 +125,103 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
 // *INDENT-ON*
 
 template <typename T>
-using CLFFT1DFixture = FFTValidationFixture<CLTensor, CLAccessor, CLFFT1D, T>;
+using CLFFT1DFixture = FFTValidationFixture<CLTensor, CLAccessor, CLFFT1D, FFT1DInfo, T>;
 
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLFFT1DFixture<float>, framework::DatasetMode::ALL, combine(shapes, framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFFT1DFixture<float>, framework::DatasetMode::ALL, combine(shapes_1d, framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
-    validate(CLAccessor(_target), _reference, RelativeTolerance<float>(0.1f), 0.05f);
+    validate(CLAccessor(_target), _reference, tolerance_f32, tolerance_num);
 }
 TEST_SUITE_END() // FP32
 TEST_SUITE_END() // Float
 
 TEST_SUITE_END() // FFT1D
+
+TEST_SUITE(FFT2D)
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(shapes_2d, data_types),
+               shape, data_type)
+{
+    // Create tensors
+    CLTensor src = create_tensor<CLTensor>(shape, data_type, 2);
+    CLTensor dst = create_tensor<CLTensor>(shape, data_type, 2);
+
+    ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    CLFFT2D fft2d;
+    fft2d.configure(&src, &dst, FFT2DInfo());
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(src.info()->valid_region(), valid_region);
+    validate(dst.info()->valid_region(), valid_region);
+
+    // Validate padding
+    validate(src.info()->padding(), PaddingSize());
+    validate(dst.info()->padding(), PaddingSize());
+}
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(
+        framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F32), // Mismatching data types
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F32), // Mismatching shapes
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 3, DataType::F32), // Invalid channels
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32), // Undecomposable FFT
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F32),
+        }),
+        framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F16),
+                                                TensorInfo(TensorShape(16U, 25U, 2U), 2, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F32),
+        })),
+        framework::dataset::make("Expected", { false, false, false, false, true })),
+               input_info, output_info, expected)
+{
+    const Status s = CLFFT2D::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), FFT2DInfo());
+    ARM_COMPUTE_EXPECT(bool(s) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+template <typename T>
+using CLFFT2DFixture = FFTValidationFixture<CLTensor, CLAccessor, CLFFT2D, FFT2DInfo, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFFT2DFixture<float>, framework::DatasetMode::ALL, combine(shapes_2d, framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, tolerance_num);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // FFT2D
+
+TEST_SUITE(FFTConvolutionLayer)
+
+template <typename T>
+using CLFFTConvolutionLayerFixture = FFTConvolutionValidationFixture<CLTensor, CLAccessor, CLFFTConvolutionLayer, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFFTConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFFTConvolutionLayerDataset(),
+                                                                                                                 framework::dataset::make("DataType", DataType::F32)),
+                                                                                                                 framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
+                                                                                                                 ActivationFunctionsSmallDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, tolerance_num);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // FFTConvolutionLayer
+
 TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/CL/ReductionOperation.cpp b/tests/validation/CL/ReductionOperation.cpp
index c8474e97e6..79308c8229 100644
--- a/tests/validation/CL/ReductionOperation.cpp
+++ b/tests/validation/CL/ReductionOperation.cpp
@@ -63,7 +63,7 @@ TEST_SUITE(ReductionOperation)
 // clang-format off
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
     framework::dataset::make("InputInfo",          { TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Mismatching data type input/output
-                                                     TensorInfo(TensorShape(128U, 64U), 2, DataType::F32), // Number of Input channels != 1
+                                                     TensorInfo(TensorShape(128U, 64U), 3, DataType::F32), // Number of Input channels != 1
                                                      TensorInfo(TensorShape(128U, 64U), 1, DataType::S16), // DataType != QASYMM8/F16/F32
                                                      TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Axis >= num_max_dimensions
                                                      TensorInfo(TensorShape(128U, 64U), 1, DataType::QASYMM8), // Axis == 0 and SUM_SQUARE and QASYMM8
diff --git a/tests/validation/fixtures/ConvolutionLayerFixture.h b/tests/validation/fixtures/ConvolutionLayerFixture.h
index 795b9de6cd..52fa8da60b 100644
--- a/tests/validation/fixtures/ConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/ConvolutionLayerFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/tests/validation/fixtures/FFTFixture.h b/tests/validation/fixtures/FFTFixture.h
index 8e3c01eaff..1aaa5965b2 100644
--- a/tests/validation/fixtures/FFTFixture.h
+++ b/tests/validation/fixtures/FFTFixture.h
@@ -31,6 +31,8 @@
 #include "tests/IAccessor.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Fixture.h"
+#include "tests/validation/reference/ActivationLayer.h"
+#include "tests/validation/reference/ConvolutionLayer.h"
 #include "tests/validation/reference/DFT.h"
 
 #include <random>
@@ -41,7 +43,7 @@ namespace test
 {
 namespace validation
 {
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+template <typename TensorType, typename AccessorType, typename FunctionType, typename InfoType, typename T>
 class FFTValidationFixture : public framework::Fixture
 {
 public:
@@ -68,8 +70,8 @@ protected:
         TensorType dst = create_tensor<TensorType>(shape, data_type, 2);
 
         // Create and configure function
-        FunctionType fft1d;
-        fft1d.configure(&src, &dst, FFT1DInfo());
+        FunctionType fft;
+        fft.configure(&src, &dst, InfoType());
 
         ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -85,7 +87,7 @@ protected:
         fill(AccessorType(src));
 
         // Compute function
-        fft1d.run();
+        fft.run();
 
         return dst;
     }
@@ -97,12 +99,138 @@ protected:
 
         // Fill reference
         fill(src);
+        if(std::is_same<InfoType, FFT1DInfo>::value)
+        {
+            return reference::dft_1d(src, reference::FFTDirection::Forward);
+        }
+        else
+        {
+            return reference::dft_2d(src, reference::FFTDirection::Forward);
+        }
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class FFTConvolutionValidationGenericFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation,
+               DataType data_type, DataLayout data_layout, ActivationLayerInfo act_info)
+    {
+        _data_type   = data_type;
+        _data_layout = data_layout;
+
+        _target    = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, dilation, act_info);
+        _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, dilation, act_info);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        switch(tensor.data_type())
+        {
+            case DataType::F32:
+            {
+                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(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, TensorShape output_shape, const PadStrideInfo &info,
+                              const Size2D &dilation, const ActivationLayerInfo act_info)
+    {
+        ARM_COMPUTE_UNUSED(dilation);
+        ARM_COMPUTE_ERROR_ON((input_shape[2] % weights_shape[2]) != 0);
+
+        if(_data_layout == DataLayout::NHWC)
+        {
+            permute(input_shape, PermutationVector(2U, 0U, 1U));
+            permute(weights_shape, PermutationVector(2U, 0U, 1U));
+            permute(output_shape, PermutationVector(2U, 0U, 1U));
+        }
+
+        // Create tensors
+        TensorType src     = create_tensor<TensorType>(input_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType weights = create_tensor<TensorType>(weights_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType bias    = create_tensor<TensorType>(bias_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType dst     = create_tensor<TensorType>(output_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+
+        // Create and configure function
+        FunctionType conv;
+        conv.configure(&src, &weights, &bias, &dst, info, act_info);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        weights.allocator()->allocate();
+        bias.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src), 0);
+        fill(AccessorType(weights), 1);
+        fill(AccessorType(bias), 2);
+
+        // Compute convolution function
+        conv.run();
+
+        return dst;
+    }
 
-        return reference::dft_1d(src, reference::FFTDirection::Forward);
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info,
+                                      const Size2D &dilation, const ActivationLayerInfo act_info)
+    {
+        ARM_COMPUTE_ERROR_ON((input_shape[2] % weights_shape[2]) != 0);
+
+        // Create reference
+        SimpleTensor<T> src{ input_shape, _data_type, 1 };
+        SimpleTensor<T> weights{ weights_shape, _data_type, 1 };
+        SimpleTensor<T> bias{ bias_shape, _data_type, 1 };
+
+        // Fill reference
+        fill(src, 0);
+        fill(weights, 1);
+        fill(bias, 2);
+
+        return (act_info.enabled()) ? reference::activation_layer<T>(reference::convolution_layer<T>(src, weights, bias, output_shape, info, dilation), act_info) : reference::convolution_layer<T>(src,
+                weights, bias, output_shape, info, dilation);
     }
 
     TensorType      _target{};
     SimpleTensor<T> _reference{};
+    DataType        _data_type{};
+    DataLayout      _data_layout{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class FFTConvolutionValidationFixture : public FFTConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation,
+               DataType data_type, DataLayout data_layout, ActivationLayerInfo act_info)
+    {
+        FFTConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation,
+                                                                                                 data_type, data_layout, act_info);
+    }
 };
 } // namespace validation
 } // namespace test
author	Georgios Pinitas <georgios.pinitas@arm.com>	2019-03-26 17:23:28 +0000
committer	Gian Marco Iodice <gianmarco.iodice@arm.com>	2019-04-11 09:34:26 +0000
commit	8be9148814b88e5b0cabd5a4d2b1f4ff470a8c1c (patch)
tree	760658b8c7b8917379467bd3fc119a5502faa850
parent	a50e702289af66944e860eafc7f3b32f6c5f30be (diff)
download	ComputeLibrary-8be9148814b88e5b0cabd5a4d2b1f4ff470a8c1c.tar.gz