COMPMID-2387: Add support for NEMeanStdDevNormalizationLayer

Change-Id: If5a9558b78f9c696dfc72633f73e9e21e5c2970f
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1351
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

7 files changed, 528 insertions, 0 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 119f9b8702..206ff06f8a 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -102,6 +102,7 @@
 #include "arm_compute/core/NEON/kernels/NELocallyConnectedMatrixMultiplyKernel.h"
 #include "arm_compute/core/NEON/kernels/NEMagnitudePhaseKernel.h"
 #include "arm_compute/core/NEON/kernels/NEMeanStdDevKernel.h"
+#include "arm_compute/core/NEON/kernels/NEMeanStdDevNormalizationKernel.h"
 #include "arm_compute/core/NEON/kernels/NEMedian3x3Kernel.h"
 #include "arm_compute/core/NEON/kernels/NEMemsetKernel.h"
 #include "arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEMeanStdDevNormalizationKernel.h b/arm_compute/core/NEON/kernels/NEMeanStdDevNormalizationKernel.h
new file mode 100644
index 0000000000..ba5410ae16
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEMeanStdDevNormalizationKernel.h
@@ -0,0 +1,98 @@
+/*
+ * 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_NEMEANSTDDEVNORMALIZATIONKERNEL_H__
+#define __ARM_COMPUTE_NEMEANSTDDEVNORMALIZATIONKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#include <arm_fp16.h>
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface for the kernel to normalize the input 2D tensor across the first dimension with respect to mean and standard deviation of the same dimension. */
+class NEMeanStdDevNormalizationKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEMeanStdDevNormalizationKernel";
+    }
+    /** Default constructor */
+    NEMeanStdDevNormalizationKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEMeanStdDevNormalizationKernel(const NEMeanStdDevNormalizationKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEMeanStdDevNormalizationKernel &operator=(const NEMeanStdDevNormalizationKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEMeanStdDevNormalizationKernel(NEMeanStdDevNormalizationKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEMeanStdDevNormalizationKernel &operator=(NEMeanStdDevNormalizationKernel &&) = default;
+    /** Default destructor */
+    ~NEMeanStdDevNormalizationKernel() = default;
+    /** Initialise the kernel's input and outputs.
+     *
+     * @note If the output tensor is a nullptr, the normalization will be performed in-place.
+     *
+     * @param[in, out] input   Source tensor with 2 dimensions. In case of @p output tensor = nullptr,
+     *                         this tensor will store the result of the normalization. Data types supported: F16/F32.
+     * @param[out]     output  (Optional) Destination tensor. It can be nullptr in case of in-place computation. Data type supported: same as @p input
+     * @param[in]      epsilon (Optional) Small float to avoid division by zero in case of zero standard deviation. Defaults to 1e-8.
+     */
+    void configure(ITensor *input, ITensor *output = nullptr, float epsilon = 1e-8f);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEMeanStdDevNormalizationKernel
+     *
+     * @param[in] input   Source tensor info with 2 dimensions. In case of @p output tensor info = nullptr,
+     *                    this tensor will store the result of the normalization. Data types supported: F16/F32.
+     * @param[in] output  (Optional) Destination tensor info. It can be nullptr in case of in-place computation. Data type supported: same as @p input
+     * @param[in] epsilon (Optional) Small float to avoid division by zero in case of zero standard deviation. Defaults to 1e-8.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output = nullptr, float epsilon = 1e-8f);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    /** Normalizes the input with respect to mean and standard deviation.
+     *
+     * @param[in] window Region on which to execute the kernel.
+     */
+    template <typename ScalarType, int size>
+    void mean_stddev_normalization(const Window &window);
+
+    ITensor *_input;
+    ITensor *_output;
+    float    _epsilon;
+
+    using MeanStdDevNormFunction = void (NEMeanStdDevNormalizationKernel::*)(const Window &window);
+
+    MeanStdDevNormFunction _func;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NEMEANSTDDEVNORMALIZATIONKERNEL_H__ */
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 94607364b3..d44afcbb0f 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -100,6 +100,7 @@
 #include "arm_compute/runtime/NEON/functions/NELocallyConnectedLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEMagnitude.h"
 #include "arm_compute/runtime/NEON/functions/NEMeanStdDev.h"
+#include "arm_compute/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEMedian3x3.h"
 #include "arm_compute/runtime/NEON/functions/NEMinMaxLocation.h"
 #include "arm_compute/runtime/NEON/functions/NENonLinearFilter.h"
diff --git a/arm_compute/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.h b/arm_compute/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.h
new file mode 100644
index 0000000000..1a482bbebd
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.h
@@ -0,0 +1,59 @@
+/*
+ * 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_NEMEANSTDDEVNORMALIZATIONLAYER_H__
+#define __ARM_COMPUTE_NEMEANSTDDEVNORMALIZATIONLAYER_H__
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/INESimpleFunctionNoBorder.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Basic function to execute mean and standard deviation normalization by calling @ref NEMeanStdDevNormalizationKernel */
+class NEMeanStdDevNormalizationLayer : public INESimpleFunctionNoBorder
+{
+public:
+    /** Initialise the function's input and outputs.
+     *
+     * @note If the output tensor is a nullptr, the normalization will be performed in-place.
+     *
+     * @param[in, out] input   Input tensor with 2 dimensions. Data types supported: F16/F32.
+     * @param[out]     output  (Optional) Destination tensor. It can be nullptr in case of in-place computation. Data type supported: same as @p input
+     * @param[in]      epsilon (Optional) Small float to avoid division by zero in case of zero standard deviation. Defaults to 1e-8.
+     */
+    void configure(ITensor *input, ITensor *output = nullptr, float epsilon = 1e-8f);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEMeanStdDevNormalizationKernel
+     *
+     * @param[in] input   Source tensor info with 2 dimensions. In case of @p output tensor info = nullptr,
+     *                    this tensor will store the result of the normalization. Data types supported: F16/F32.
+     * @param[in] output  (Optional) Destination tensor info. It can be nullptr in case of in-place computation. Data type supported: same as @p input
+     * @param[in] epsilon (Optional) Small float to avoid division by zero in case of zero standard deviation. Defaults to 1e-8.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output = nullptr, float epsilon = 1e-8f);
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_NEMEANSTDDEVNORMALIZATIONLAYER_H__ */
diff --git a/src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp b/src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp
new file mode 100644
index 0000000000..57380b4546
--- /dev/null
+++ b/src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp
@@ -0,0 +1,203 @@
+/*
+ * 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/NEON/kernels/NEMeanStdDevNormalizationKernel.h"
+
+#include "arm_compute/core/CPP/Validate.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/NEMath.h"
+#include "arm_compute/core/NEON/wrapper/wrapper.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, float epsilon)
+{
+    ARM_COMPUTE_UNUSED(epsilon);
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 2, "Input tensor cannot have more than 2 dimensions");
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+
+    // 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{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    if(output != nullptr)
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+        // Output auto inizialitation if not yet initialized
+        auto_init_if_empty(*output, *input);
+    }
+
+    // This kernel doesn't need padding. A left-over for loop on dimension X, we cannot have any read or write out of memory
+    // For this reason num_elems_processed_per_iteration is set to 1
+    Window win = calculate_max_window(*input, Steps());
+    if(output != nullptr)
+    {
+        output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+    }
+
+    return std::make_pair(Status{}, win);
+}
+} // namespace
+
+template <typename ScalarType, int size>
+void NEMeanStdDevNormalizationKernel::mean_stddev_normalization(const Window &window)
+{
+    using ExactTagType = typename wrapper::traits::neon_vector<ScalarType, size>::tag_type;
+
+    // Set build options
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x  = size;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Iterator input(_input, win);
+    Iterator output(_output, win);
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        int  x       = window_start_x;
+        auto in_ptr  = reinterpret_cast<const ScalarType *>(input.ptr());
+        auto out_ptr = reinterpret_cast<ScalarType *>(output.ptr());
+
+        auto sum_vec    = wrapper::vdup_n(static_cast<ScalarType>(0.f), ExactTagType{});
+        auto sum_sq_vec = wrapper::vdup_n(static_cast<ScalarType>(0.f), ExactTagType{});
+
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            auto data  = wrapper::vloadq(in_ptr + x);
+            sum_vec    = wrapper::vadd(sum_vec, data);
+            sum_sq_vec = wrapper::vadd(sum_sq_vec, wrapper::vmul(data, data));
+        }
+
+        auto sum_carry_res    = wrapper::vpadd(wrapper::vgethigh(sum_vec), wrapper::vgetlow(sum_vec));
+        auto sum_sq_carry_res = wrapper::vpadd(wrapper::vgethigh(sum_sq_vec), wrapper::vgetlow(sum_sq_vec));
+        for(int i = 0; i < size / 4; ++i)
+        {
+            sum_carry_res    = wrapper::vpadd(sum_carry_res, sum_carry_res);
+            sum_sq_carry_res = wrapper::vpadd(sum_sq_carry_res, sum_sq_carry_res);
+        }
+
+        auto sum    = wrapper::vgetlane(sum_carry_res, 0);
+        auto sum_sq = wrapper::vgetlane(sum_sq_carry_res, 0);
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            ScalarType data = *(in_ptr + x);
+            sum += data;
+            sum_sq += data * data;
+        }
+
+        ScalarType mean       = sum / _input->info()->dimension(0);
+        ScalarType var        = (sum_sq / _input->info()->dimension(0)) - (mean * mean);
+        ScalarType stddev_inv = 1.f / sqrt(var + _epsilon);
+
+        auto mean_vec       = wrapper::vdup_n(mean, ExactTagType{});
+        auto stddev_inv_vec = wrapper::vdup_n(stddev_inv, ExactTagType{});
+        for(x = window_start_x; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            auto data = wrapper::vloadq(in_ptr + x);
+            auto res  = wrapper::vmul(wrapper::vsub(data, mean_vec), stddev_inv_vec);
+            // Store results
+            wrapper::vstore(out_ptr + x, res);
+        }
+        for(; x < window_end_x; ++x)
+        {
+            *(out_ptr + x) = (*(in_ptr + x) - mean) * stddev_inv;
+        }
+    },
+    input, output);
+}
+
+NEMeanStdDevNormalizationKernel::NEMeanStdDevNormalizationKernel()
+    : _input(nullptr), _output(nullptr), _epsilon(1e-8f), _func(nullptr)
+{
+}
+
+void NEMeanStdDevNormalizationKernel::configure(ITensor *input, ITensor *output, float epsilon)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+
+    ARM_COMPUTE_ERROR_THROW_ON(NEMeanStdDevNormalizationKernel::validate(input->info(), (output != nullptr) ? output->info() : nullptr, epsilon));
+
+    _input   = input;
+    _output  = (output == nullptr) ? input : output;
+    _epsilon = epsilon;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), (output == nullptr) ? nullptr : output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICPPKernel::configure(win_config.second);
+
+    // Configure function to run based on different data types
+    const DataType data_type = input->info()->data_type();
+    switch(data_type)
+    {
+        case DataType::F32:
+            _func = &NEMeanStdDevNormalizationKernel::mean_stddev_normalization<float, 4>;
+            break;
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        case DataType::F16:
+            _func = &NEMeanStdDevNormalizationKernel::mean_stddev_normalization<float16_t, 8>;
+            break;
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        default:
+            ARM_COMPUTE_ERROR("Not Supported");
+            break;
+    }
+}
+
+Status NEMeanStdDevNormalizationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, float epsilon)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, epsilon));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (output != nullptr) ? output->clone().get() : nullptr).first);
+    return Status{};
+}
+
+void NEMeanStdDevNormalizationKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON(_func == nullptr);
+
+    (this->*_func)(window);
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.cpp b/src/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.cpp
new file mode 100644
index 0000000000..3ecebb42bc
--- /dev/null
+++ b/src/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.cpp
@@ -0,0 +1,42 @@
+/*
+ * 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/NEON/functions/NEMeanStdDevNormalizationLayer.h"
+
+#include "arm_compute/core/NEON/kernels/NEMeanStdDevNormalizationKernel.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+void NEMeanStdDevNormalizationLayer::configure(ITensor *input, ITensor *output, float epsilon)
+{
+    auto k = arm_compute::support::cpp14::make_unique<NEMeanStdDevNormalizationKernel>();
+    k->configure(input, output, epsilon);
+    _kernel = std::move(k);
+}
+
+Status NEMeanStdDevNormalizationLayer::validate(const ITensorInfo *input, const ITensorInfo *output, float epsilon)
+{
+    return NEMeanStdDevNormalizationKernel::validate(input, output, epsilon);
+}
+} // namespace arm_compute
diff --git a/tests/validation/NEON/MeanStdDevNormalizationLayer.cpp b/tests/validation/NEON/MeanStdDevNormalizationLayer.cpp
new file mode 100644
index 0000000000..132700929c
--- /dev/null
+++ b/tests/validation/NEON/MeanStdDevNormalizationLayer.cpp
@@ -0,0 +1,124 @@
+/*
+ * 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/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/NormalizationTypesDataset.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/MeanStdDevNormalizationLayerFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+/** Tolerance for float operations */
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+RelativeTolerance<half> tolerance_f16(half(0.2f));
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+RelativeTolerance<float> tolerance_f32(1e-8f);
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(MeanStdDevNormalizationLayer)
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(
+               framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), // Mismatching data type input/output
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), // Mismatching shapes
+                                                       TensorInfo(TensorShape(32U, 13U), 1, DataType::F32),
+                                                     }),
+               framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U), 1, DataType::F16),
+                                                       TensorInfo(TensorShape(27U, 11U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U), 1, DataType::F32),
+                                                     })),
+               framework::dataset::make("Expected", { false, false, true })),
+               input_info, output_info, expected)
+{
+    ARM_COMPUTE_EXPECT(bool(NEMeanStdDevNormalizationLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+template <typename T>
+using NEMeanStdDevNormalizationLayerFixture = MeanStdDevNormalizationLayerValidationFixture<Tensor, Accessor, NEMeanStdDevNormalizationLayer, T>;
+
+TEST_SUITE(Float)
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEMeanStdDevNormalizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small2DShapes(),
+                       framework::dataset::make("DataType", DataType::F16)),
+                       framework::dataset::make("InPlace", { false, true })),
+                       framework::dataset::make("Epsilon", { 1e-8 })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEMeanStdDevNormalizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::Large2DShapes(),
+                                                                                                                       framework::dataset::make("DataType", DataType::F16)),
+                                                                                                                       framework::dataset::make("InPlace", { false, true })),
+                                                                                                                       framework::dataset::make("Epsilon", { 1e-8 })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END() // FP16
+#endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEMeanStdDevNormalizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small2DShapes(),
+                       framework::dataset::make("DataType", DataType::F32)),
+                       framework::dataset::make("InPlace", { false, true })),
+                       framework::dataset::make("Epsilon", { 1e-8 })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEMeanStdDevNormalizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::Large2DShapes(),
+                                                                                                                        framework::dataset::make("DataType", DataType::F32)),
+                                                                                                                        framework::dataset::make("InPlace", { false, true })),
+                                                                                                                        framework::dataset::make("Epsilon", { 1e-8 })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+
+TEST_SUITE_END() // MeanStdNormalizationLayer
+TEST_SUITE_END() // NEON
+} // namespace validation
+} // namespace test
+} // namespace arm_compute