COMPMID-1767: NEON: Implement Where/Select

Change-Id: If8a1ab6d6a029a5c547b726e0692eecef9a2e97d
Reviewed-on: https://review.mlplatform.org/415
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

10 files changed, 789 insertions, 0 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 57a1d4d52c..e859519d94 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -110,6 +110,7 @@
 #include "arm_compute/core/NEON/kernels/NEReshapeLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEScaleKernel.h"
 #include "arm_compute/core/NEON/kernels/NEScharr3x3Kernel.h"
+#include "arm_compute/core/NEON/kernels/NESelectKernel.h"
 #include "arm_compute/core/NEON/kernels/NESobel3x3Kernel.h"
 #include "arm_compute/core/NEON/kernels/NESobel5x5Kernel.h"
 #include "arm_compute/core/NEON/kernels/NESobel7x7Kernel.h"
diff --git a/arm_compute/core/NEON/kernels/NESelectKernel.h b/arm_compute/core/NEON/kernels/NESelectKernel.h
new file mode 100644
index 0000000000..215dc2f5ee
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NESelectKernel.h
@@ -0,0 +1,102 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INNEUDING 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 NEAIM, 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_NESELECTKERNEL_H__
+#define __ARM_COMPUTE_NESELECTKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface for the select kernel
+ *
+ * Select is computed by:
+ * @f[ output(i) = condition(i) ? x(i) : y(i) @f]
+ *
+ */
+class NESelectKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NESelectKernel";
+    }
+    /** Default constructor */
+    NESelectKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NESelectKernel(const NESelectKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NESelectKernel &operator=(const NESelectKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NESelectKernel(NESelectKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NESelectKernel &operator=(NESelectKernel &&) = default;
+    /** Default destructor */
+    ~NESelectKernel() = default;
+
+    /** Common signature for all the specialised elementwise functions
+     *
+     * @param[in]  c      Condition input tensor. Data types supported: U8.
+     * @param[in]  x      First input tensor. Data types supported: U8/S8/U16/S16/U32/S32/F16/F32.
+     * @param[out] y      Second input tensor. Data types supported: Same as @p x
+     * @param[in]  output Output tensor. Data types supported: Same as @p x
+     */
+    void configure(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output);
+
+    /** Validate the argument passed to the kernel
+     *
+     * @param[in] c      Condition input tensor. Data types supported: U8.
+     * @param[in] x      First input tensor. Data types supported: U8/S8/U16/S16/U32/S32/F16/F32.
+     * @param[in] y      Second input tensor. Data types supported: Same as @p x
+     * @param[in] output Output tensor. Data types supported: Same as @p x.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *c, const ITensorInfo *x, const ITensorInfo *y, const ITensorInfo *output);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    /** Common signature for all the specialised select functions
+     *
+     * @param[in] c      Condition input tensor. Data types supported: U8.
+     * @param[in] x      First input tensor. Data types supported: U8/S8/U16/S16/U32/S32/F16/F32.
+     * @param[in] y      Second input tensor. Data types supported: Same as @p x
+     * @param[in] output Output tensor. Data types supported: Same as @p x.
+     */
+    using SelectFunction = void(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window);
+
+    /** Select function to use for the particular tensor types passed to configure() */
+    SelectFunction *_function;
+    const ITensor *_c;              /**< Condition tensor */
+    const ITensor *_x;              /**< Source tensor 1 */
+    const ITensor *_y;              /**< Source tensor 2 */
+    ITensor        *_output;        /**< Destination tensor */
+    bool            _has_same_rank; /**< Flag that indicates if condition tensor and other inputs have the same rank */
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NESELECTKERNEL_H__ */
diff --git a/arm_compute/core/NEON/wrapper/intrinsics/bitselect.h b/arm_compute/core/NEON/wrapper/intrinsics/bitselect.h
new file mode 100644
index 0000000000..8223f6d463
--- /dev/null
+++ b/arm_compute/core/NEON/wrapper/intrinsics/bitselect.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT SELECT 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_WRAPPER_BITSELECT_H__
+#define __ARM_COMPUTE_WRAPPER_BITSELECT_H__
+
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+namespace wrapper
+{
+#define VBITSELECT_IMPL(stype, vtype, ctype, prefix, postfix)               \
+    inline vtype vbitselect(const ctype &a, const vtype &b, const vtype &c) \
+    {                                                                       \
+        return prefix##_##postfix(a, b, c);                                 \
+    }
+
+VBITSELECT_IMPL(uint8_t, uint8x8_t, uint8x8_t, vbsl, u8)
+VBITSELECT_IMPL(int8_t, int8x8_t, uint8x8_t, vbsl, s8)
+VBITSELECT_IMPL(uint16_t, uint16x4_t, uint16x4_t, vbsl, u16)
+VBITSELECT_IMPL(int16_t, int16x4_t, uint16x4_t, vbsl, s16)
+VBITSELECT_IMPL(uint32_t, uint32x2_t, uint32x2_t, vbsl, u32)
+VBITSELECT_IMPL(int32_t, int32x2_t, uint32x2_t, vbsl, s32)
+VBITSELECT_IMPL(float32x2_t, float32x2_t, uint32x2_t, vbsl, f32)
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+VBITSELECT_IMPL(float16x4_t, float16x4_t, uint16x4_t, vbsl, f16)
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+VBITSELECT_IMPL(uint8_t, uint8x16_t, uint8x16_t, vbslq, u8)
+VBITSELECT_IMPL(int8_t, int8x16_t, uint8x16_t, vbslq, s8)
+VBITSELECT_IMPL(uint16_t, uint16x8_t, uint16x8_t, vbslq, u16)
+VBITSELECT_IMPL(int16_t, int16x8_t, uint16x8_t, vbslq, s16)
+VBITSELECT_IMPL(uint32_t, uint32x4_t, uint32x4_t, vbslq, u32)
+VBITSELECT_IMPL(int32_t, int32x4_t, uint32x4_t, vbslq, s32)
+VBITSELECT_IMPL(float32x4_t, float32x4_t, uint32x4_t, vbslq, f32)
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+VBITSELECT_IMPL(float16x8_t, float16x8_t, uint16x8_t, vbslq, f16)
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#undef VBITSELECT_IMPL
+} // namespace wrapper
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_WRAPPER_BITSELECT_H__ */
diff --git a/arm_compute/core/NEON/wrapper/intrinsics/greaterthan.h b/arm_compute/core/NEON/wrapper/intrinsics/greaterthan.h
new file mode 100644
index 0000000000..5ee7516a4e
--- /dev/null
+++ b/arm_compute/core/NEON/wrapper/intrinsics/greaterthan.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_WRAPPER_CGT_H__
+#define __ARM_COMPUTE_WRAPPER_CGT_H__
+
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+namespace wrapper
+{
+#define VCGT_IMPL(stype, vtype, rtype, prefix, postfix)       \
+    inline rtype vgreaterthan(const vtype &a, const vtype &b) \
+    {                                                         \
+        return prefix##_##postfix(a, b);                      \
+    }
+
+VCGT_IMPL(uint8_t, uint8x8_t, uint8x8_t, vcgt, u8)
+VCGT_IMPL(int8_t, int8x8_t, uint8x8_t, vcgt, s8)
+VCGT_IMPL(uint16_t, uint16x4_t, uint16x4_t, vcgt, u16)
+VCGT_IMPL(int16_t, int16x4_t, uint16x4_t, vcgt, s16)
+VCGT_IMPL(uint32_t, uint32x2_t, uint32x2_t, vcgt, u32)
+VCGT_IMPL(int32_t, int32x2_t, uint32x2_t, vcgt, s32)
+VCGT_IMPL(float32x2_t, float32x2_t, uint32x2_t, vcgt, f32)
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+VCGT_IMPL(float16x4_t, float16x4_t, uint16x4_t, vcgt, f16)
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+VCGT_IMPL(uint8_t, uint8x16_t, uint8x16_t, vcgtq, u8)
+VCGT_IMPL(int8_t, int8x16_t, uint8x16_t, vcgtq, s8)
+VCGT_IMPL(uint16_t, uint16x8_t, uint16x8_t, vcgtq, u16)
+VCGT_IMPL(int16_t, int16x8_t, uint16x8_t, vcgtq, s16)
+VCGT_IMPL(uint32_t, uint32x4_t, uint32x4_t, vcgtq, u32)
+VCGT_IMPL(int32_t, int32x4_t, uint32x4_t, vcgtq, s32)
+VCGT_IMPL(float32x4_t, float32x4_t, uint32x4_t, vcgtq, f32)
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+VCGT_IMPL(float16x8_t, float16x8_t, uint16x8_t, vcgtq, f16)
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#undef VCGT_IMPL
+} // namespace wrapper
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_WRAPPER_CGT_H__ */
diff --git a/arm_compute/core/NEON/wrapper/intrinsics/intrinsics.h b/arm_compute/core/NEON/wrapper/intrinsics/intrinsics.h
index 46c8937adc..f65ce85021 100644
--- a/arm_compute/core/NEON/wrapper/intrinsics/intrinsics.h
+++ b/arm_compute/core/NEON/wrapper/intrinsics/intrinsics.h
@@ -26,11 +26,13 @@
 
 #include "arm_compute/core/NEON/wrapper/intrinsics/add.h"
 #include "arm_compute/core/NEON/wrapper/intrinsics/and.h"
+#include "arm_compute/core/NEON/wrapper/intrinsics/bitselect.h"
 #include "arm_compute/core/NEON/wrapper/intrinsics/dup_n.h"
 #include "arm_compute/core/NEON/wrapper/intrinsics/exp.h"
 #include "arm_compute/core/NEON/wrapper/intrinsics/gethigh.h"
 #include "arm_compute/core/NEON/wrapper/intrinsics/getlane.h"
 #include "arm_compute/core/NEON/wrapper/intrinsics/getlow.h"
+#include "arm_compute/core/NEON/wrapper/intrinsics/greaterthan.h"
 #include "arm_compute/core/NEON/wrapper/intrinsics/inv.h"
 #include "arm_compute/core/NEON/wrapper/intrinsics/invsqrt.h"
 #include "arm_compute/core/NEON/wrapper/intrinsics/load.h"
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 2ffd92d541..bfecbcf65e 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -112,6 +112,7 @@
 #include "arm_compute/runtime/NEON/functions/NEReshapeLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEScale.h"
 #include "arm_compute/runtime/NEON/functions/NEScharr3x3.h"
+#include "arm_compute/runtime/NEON/functions/NESelect.h"
 #include "arm_compute/runtime/NEON/functions/NESimpleAssemblyFunction.h"
 #include "arm_compute/runtime/NEON/functions/NESobel3x3.h"
 #include "arm_compute/runtime/NEON/functions/NESobel5x5.h"
diff --git a/arm_compute/runtime/NEON/functions/NESelect.h b/arm_compute/runtime/NEON/functions/NESelect.h
new file mode 100644
index 0000000000..1529cd6963
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NESelect.h
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_NESELECT_H__
+#define __ARM_COMPUTE_NESELECT_H__
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/INESimpleFunction.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ITensor;
+
+/** Basic function to run @ref NESelect */
+class NESelect : public INESimpleFunction
+{
+public:
+    /** Initialise the kernel's inputs and output.
+     *
+     * @param[in]  c      Condition input tensor. Data types supported: U8.
+     * @param[in]  x      First input tensor. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
+     * @param[in]  y      Second input tensor. Data types supported: Same as @p x
+     * @param[out] output Output tensor. Data types supported: Same as @p x.
+     */
+    void configure(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output);
+    /** Static function to check if given info will lead to a valid configuration of @ref NESelect
+     *
+     * @param[in] c      Condition input tensor. Data types supported: U8.
+     * @param[in] x      First input tensor. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32.
+     * @param[in] y      Second input tensor. Data types supported: Same as @p x
+     * @param[in] output Output tensor. Data types supported: Same as @p x.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *c, const ITensorInfo *x, const ITensorInfo *y, const ITensorInfo *output);
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NESELECT_H__ */
diff --git a/src/core/NEON/kernels/NESelectKernel.cpp b/src/core/NEON/kernels/NESelectKernel.cpp
new file mode 100644
index 0000000000..0c134c00ed
--- /dev/null
+++ b/src/core/NEON/kernels/NESelectKernel.cpp
@@ -0,0 +1,264 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INNEUDING 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 NEAIM, 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/NESelectKernel.h"
+
+#include "arm_compute/core/CPP/Validate.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/ITensor.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/Validate.h"
+#include "utils/TypePrinter.h"
+
+#include <arm_neon.h>
+#include <map>
+#include <string>
+
+namespace arm_compute
+{
+namespace
+{
+template <typename ScalarType, typename VectorType>
+void select_op(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
+               const int window_step_x, const int window_start_x, const int window_end_x, const int limit, VectorType (*condition_conversion)(const uint8_t *))
+{
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator condition(cond, win);
+    Iterator input1(in1, win);
+    Iterator input2(in2, win);
+    Iterator output(out, win);
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        auto       output_ptr    = reinterpret_cast<ScalarType *>(output.ptr());
+        const auto condition_ptr = reinterpret_cast<const uint8_t *>(condition.ptr());
+        const auto input1_ptr    = reinterpret_cast<const ScalarType *>(input1.ptr());
+        const auto input2_ptr    = reinterpret_cast<const ScalarType *>(input2.ptr());
+
+        int x = window_start_x;
+        for(; x <= limit; x += window_step_x)
+        {
+            const auto c = (*condition_conversion)(condition_ptr + x);
+            const auto a = wrapper::vloadq(input1_ptr + x);
+            const auto b = wrapper::vloadq(input2_ptr + x);
+            wrapper::vstore(output_ptr + x, wrapper::vbitselect(c, a, b));
+        }
+        for(; x < window_end_x; ++x)
+        {
+            const auto c      = *(condition_ptr + x);
+            const auto a      = *(input1_ptr + x);
+            const auto b      = *(input2_ptr + x);
+            *(output_ptr + x) = static_cast<bool>(c) ? a : b;
+        }
+    },
+    condition, input1, input2, output);
+}
+
+template <typename ScalarType, typename VectorType>
+void select_op_8(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    const auto window_step_x  = 16 / sizeof(ScalarType);
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr)
+    {
+        static const auto zero = wrapper::vdup_n(static_cast<uint8_t>(0), arm_compute::wrapper::traits::vector_128_tag());
+        return wrapper::vgreaterthan(wrapper::vloadq(condition_ptr), zero);
+    });
+}
+
+template <typename ScalarType, typename VectorType>
+void select_op_16(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    const auto window_step_x  = 16 / sizeof(ScalarType);
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr)
+    {
+        static const auto zero = wrapper::vdup_n(static_cast<uint16_t>(0), arm_compute::wrapper::traits::vector_128_tag());
+        return wrapper::vgreaterthan(wrapper::vmovl(wrapper::vload(condition_ptr)), zero);
+    });
+}
+
+template <typename ScalarType, typename VectorType>
+void select_op_32(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    const auto window_step_x  = 16 / sizeof(ScalarType);
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr)
+    {
+        static const auto zero = wrapper::vdup_n(static_cast<uint32_t>(0), arm_compute::wrapper::traits::vector_128_tag());
+        return wrapper::vgreaterthan(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vload(condition_ptr)))), zero);
+    });
+}
+
+template <typename ScalarType>
+void select_op_not_same_rank(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(window);
+
+    auto       output_ptr    = reinterpret_cast<ScalarType *>(out->buffer());
+    const auto condition_ptr = reinterpret_cast<const uint8_t *>(cond->buffer());
+    const auto input1_ptr    = reinterpret_cast<const ScalarType *>(in1->buffer());
+    const auto input2_ptr    = reinterpret_cast<const ScalarType *>(in2->buffer());
+
+    const int outer_size = cond->info()->total_size() / cond->info()->element_size();
+    const int inner_size = (in1->info()->total_size() / in1->info()->element_size()) / outer_size;
+    int       offset     = 0;
+    const int step       = 16 / in1->info()->element_size();
+
+    for(int i = 0; i < outer_size; ++i)
+    {
+        int        x         = offset;
+        const auto input_ptr = static_cast<bool>(*(condition_ptr + i)) ? input1_ptr : input2_ptr;
+        for(; x <= offset + inner_size - step; x += step)
+        {
+            wrapper::vstore(output_ptr + x, wrapper::vloadq(input_ptr + x));
+        }
+        if(x <= offset + inner_size - (step / 2))
+        {
+            wrapper::vstore(output_ptr + x, wrapper::vload(input_ptr + x));
+            x += step / 2;
+        }
+        for(; x < offset + inner_size; ++x)
+        {
+            *(output_ptr + x) = *(input_ptr + x);
+        }
+        offset += inner_size;
+    }
+}
+} // namespace
+
+NESelectKernel::NESelectKernel()
+    : _function(nullptr), _c(nullptr), _x(nullptr), _y(nullptr), _output(nullptr), _has_same_rank(false)
+{
+}
+
+void NESelectKernel::configure(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(c, x, y, output);
+
+    // Auto initialize output if not initialized
+    auto_init_if_empty(*output->info(), x->info()->tensor_shape(), 1, x->info()->data_type());
+    ARM_COMPUTE_ERROR_THROW_ON(validate(c->info(), x->info(), y->info(), output->info()));
+
+    _c             = c;
+    _x             = x;
+    _y             = y;
+    _output        = output;
+    _has_same_rank = (c->info()->tensor_shape().num_dimensions() == x->info()->tensor_shape().num_dimensions());
+
+    std::string function_to_call("op_");
+    function_to_call += string_from_data_type(x->info()->data_type());
+
+    static std::map<std::string, SelectFunction *> map_function;
+
+    if(_has_same_rank)
+    {
+        map_function =
+        {
+            { "op_S8", &select_op_8<int8_t, uint8x16_t> },
+            { "op_S16", &select_op_16<int16_t, uint16x8_t> },
+            { "op_S32", &select_op_32<int32_t, uint32x4_t> },
+            { "op_U8", &select_op_8<uint8_t, uint8x16_t> },
+            { "op_U16", &select_op_16<uint16_t, uint16x8_t> },
+            { "op_U32", &select_op_32<uint32_t, uint32x4_t> },
+            { "op_F32", &select_op_32<float, uint32x4_t> }
+        };
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        map_function["op_F16"] = &select_op_16<float16_t, uint16x8_t>;
+#endif /* ARM_COMPUTE_AARCH64_V8_2 */
+    }
+    else
+    {
+        map_function =
+        {
+            { "op_S8", &select_op_not_same_rank<int8_t> },
+            { "op_S16", &select_op_not_same_rank<int16_t> },
+            { "op_S32", &select_op_not_same_rank<int32_t> },
+            { "op_U8", &select_op_not_same_rank<uint8_t> },
+            { "op_U16", &select_op_not_same_rank<uint16_t> },
+            { "op_U32", &select_op_not_same_rank<uint32_t> },
+            { "op_F32", &select_op_not_same_rank<float> }
+        };
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        map_function["op_F16"] = &select_op_not_same_rank<float16_t>;
+#endif /* ARM_COMPUTE_AARCH64_V8_2 */
+    }
+
+    auto it = map_function.find(function_to_call);
+
+    if(it != map_function.end())
+    {
+        _function = it->second;
+    }
+
+    Window win = calculate_max_window(x->info()->valid_region());
+    INEKernel::configure(win);
+}
+
+Status NESelectKernel::validate(const ITensorInfo *c, const ITensorInfo *x, const ITensorInfo *y, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(x);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(x,
+                                                         1,
+                                                         DataType::U8, DataType::S8,
+                                                         DataType::U16, DataType::S16,
+                                                         DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(x, y);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(x, y);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::U8);
+
+    const bool is_same_rank = (c->tensor_shape().num_dimensions() == x->tensor_shape().num_dimensions());
+    ARM_COMPUTE_RETURN_ERROR_ON(is_same_rank && (x->tensor_shape() != c->tensor_shape()));
+    ARM_COMPUTE_RETURN_ERROR_ON(!is_same_rank && ((c->tensor_shape().num_dimensions() > 1) || (c->tensor_shape().x() != x->tensor_shape()[x->tensor_shape().num_dimensions() - 1])));
+
+    if(output != nullptr && output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(x, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(x, output);
+    }
+
+    return Status{};
+}
+
+void NESelectKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON(_function == nullptr);
+    _function(_c, _x, _y, _output, window);
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NESelect.cpp b/src/runtime/NEON/functions/NESelect.cpp
new file mode 100644
index 0000000000..509bbaa24e
--- /dev/null
+++ b/src/runtime/NEON/functions/NESelect.cpp
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/NEON/functions/NESelect.h"
+
+#include "arm_compute/core/NEON/kernels/NESelectKernel.h"
+#include "arm_compute/core/Types.h"
+
+using namespace arm_compute;
+
+namespace arm_compute
+{
+void NESelect::configure(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output)
+{
+    auto k = arm_compute::support::cpp14::make_unique<NESelectKernel>();
+    k->configure(c, x, y, output);
+    _kernel = std::move(k);
+}
+
+Status NESelect::validate(const ITensorInfo *c, const ITensorInfo *x, const ITensorInfo *y, const ITensorInfo *output)
+{
+    return NESelectKernel::validate(c, x, y, output);
+}
+} // namespace arm_compute
diff --git a/tests/validation/NEON/Select.cpp b/tests/validation/NEON/Select.cpp
new file mode 100644
index 0000000000..9ac7a6f013
--- /dev/null
+++ b/tests/validation/NEON/Select.cpp
@@ -0,0 +1,188 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONCLCTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NESelect.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/SelectFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+auto configuration_dataset = combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()),
+                                     framework::dataset::make("has_same_rank", { false, true }));
+auto run_small_dataset = combine(datasets::SmallShapes(), framework::dataset::make("has_same_rank", { false, true }));
+auto run_large_dataset = combine(datasets::LargeShapes(), framework::dataset::make("has_same_rank", { false, true }));
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(Select)
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
+        framework::dataset::make("CInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), // Invalid condition datatype
+                                            TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid output datatype
+                                            TensorInfo(TensorShape(13U), 1, DataType::U8),          // Invalid c shape
+                                            TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Mismatching shapes
+                                            TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                            TensorInfo(TensorShape(2U), 1, DataType::U8),
+        }),
+        framework::dataset::make("XInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                           TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                           TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                           TensorInfo(TensorShape(32U, 10U, 2U), 1, DataType::F32),
+                                           TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                           TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+        })),
+        framework::dataset::make("YInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                           TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                           TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                           TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                           TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                           TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+        })),
+        framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+        })),
+        framework::dataset::make("Expected", { false, false, false, false, true, true})),
+        c_info, x_info, y_info, output_info, expected)
+{
+    Status s = NESelect::validate(&c_info.clone()->set_is_resizable(false),
+                                  &x_info.clone()->set_is_resizable(false),
+                                  &y_info.clone()->set_is_resizable(false),
+                                  &output_info.clone()->set_is_resizable(false));
+    ARM_COMPUTE_EXPECT(bool(s) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+template <typename T>
+using NESelectFixture = SelectValidationFixture<Tensor, Accessor, NESelect, T>;
+
+TEST_SUITE(Float)
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(F16)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, configuration_dataset,
+               shape, same_rank)
+{
+    const DataType dt = DataType::F16;
+
+    // Create tensors
+    Tensor ref_c = create_tensor<Tensor>(detail::select_condition_shape(shape, same_rank), DataType::U8);
+    Tensor ref_x = create_tensor<Tensor>(shape, dt);
+    Tensor ref_y = create_tensor<Tensor>(shape, dt);
+    Tensor dst   = create_tensor<Tensor>(shape, dt);
+
+    // Create and Configure function
+    NESelect select;
+    select.configure(&ref_c, &ref_x, &ref_y, &dst);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(dst.info()->valid_region(), valid_region);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall,
+                       NESelectFixture<half>,
+                       framework::DatasetMode::PRECOMMIT,
+                       combine(run_small_dataset, framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+                       NESelectFixture<half>,
+                       framework::DatasetMode::NIGHTLY,
+                       combine(run_large_dataset, framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // F16
+#endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
+TEST_SUITE(FP32)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, configuration_dataset,
+               shape, same_rank)
+{
+    const DataType dt = DataType::F32;
+
+    // Create tensors
+    Tensor ref_c = create_tensor<Tensor>(detail::select_condition_shape(shape, same_rank), DataType::U8);
+    Tensor ref_x = create_tensor<Tensor>(shape, dt);
+    Tensor ref_y = create_tensor<Tensor>(shape, dt);
+    Tensor dst   = create_tensor<Tensor>(shape, dt);
+
+    // Create and Configure function
+    NESelect select;
+    select.configure(&ref_c, &ref_x, &ref_y, &dst);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(dst.info()->valid_region(), valid_region);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall,
+                       NESelectFixture<float>,
+                       framework::DatasetMode::PRECOMMIT,
+                       combine(run_small_dataset, framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+                       NESelectFixture<float>,
+                       framework::DatasetMode::NIGHTLY,
+                       combine(run_large_dataset, framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+
+TEST_SUITE_END() // Select
+TEST_SUITE_END() // NEON
+} // namespace validation
+} // namespace test
+} // namespace arm_compute