Add SVE support for non-quantized elementwise operations

SVE kernels are added to all previously supported arithmetic
and comparison operations with exception of S16 arithmetic
operations due to complexity of widening and narrowing of
integer vectors.

Partially implements: COMPMID-3872

Change-Id: Ic433eb7227dfcfd0d1429f18acebec2d934ca8bd
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4778
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 461 insertions, 3 deletions

diff --git a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
index 412ae247cb..29ae9037af 100644
--- a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
+++ b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,6 +29,7 @@
 #include "src/core/NEON/NEAsymm.h"
 #include "src/core/NEON/NEFixedPoint.h"
 #include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/SVE/kernels/elementwise/impl/elementwise_list.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
 
@@ -1135,14 +1136,23 @@ configure_arithm_func(const ITensorInfo *input1, const ITensorInfo *input2, ITen
 {
     static std::map<std::string, NEElementwiseOperationKernel::ElementwiseFunction *> map_function =
     {
+#if defined(__ARM_FEATURE_SVE)
+        { "op_F32_F32_F32", &arm_compute::cpu::sve::elementwise_arithmetic_op<op, float32_t> },
+        { "op_S32_S32_S32", &arm_compute::cpu::sve::elementwise_arithmetic_op<op, int32_t> },
+#else  /* defined(__ARM_FEATURE_SVE) */
         { "op_F32_F32_F32", &elementwise_arithm_op<op, typename wrapper::traits::neon_vector<float, 4>> },
-        { "op_S16_S16_S16", &elementwise_arithm_op<op, typename wrapper::traits::neon_vector<int16_t, 8>> },
         { "op_S32_S32_S32", &elementwise_arithm_op<op, typename wrapper::traits::neon_vector<int32_t, 4>> },
+#endif /* defined(__ARM_FEATURE_SVE) */
+        { "op_S16_S16_S16", &elementwise_arithm_op<op, typename wrapper::traits::neon_vector<int16_t, 8>> },
         { "op_QASYMM8_QASYMM8_QASYMM8", &elementwise_arithm_op_quantized<op> },
         { "op_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &elementwise_arithm_op_quantized_signed<op> }
     };
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if defined(__ARM_FEATURE_SVE)
+    map_function["op_F16_F16_F16"] = &arm_compute::cpu::sve::elementwise_arithmetic_op<op, float16_t>;
+#else  /* defined(__ARM_FEATURE_SVE) */
     map_function["op_F16_F16_F16"] = &elementwise_arithm_op<op, typename wrapper::traits::neon_vector<float16_t, 8>>;
+#endif /* defined(__ARM_FEATURE_SVE) */
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
     return configure_func(input1, input2, output, map_function);
@@ -1154,15 +1164,26 @@ configure_comp_func(const ITensorInfo *input1, const ITensorInfo *input2, ITenso
 {
     static std::map<std::string, NEElementwiseOperationKernel::ElementwiseFunction *> map_function =
     {
+#if defined(__ARM_FEATURE_SVE)
+        { "op_U8_U8_U8", &arm_compute::cpu::sve::elementwise_comparison_op<op, uint8_t> },
+        { "op_F32_F32_U8", &arm_compute::cpu::sve::elementwise_comparison_op<op, float> },
+        { "op_S16_S16_U8", &arm_compute::cpu::sve::elementwise_comparison_op<op, int16_t> },
+        { "op_S32_S32_U8", &arm_compute::cpu::sve::elementwise_comparison_op<op, int32_t> },
+#else  /* defined(__ARM_FEATURE_SVE) */
         { "op_U8_U8_U8", &elementwise_comp_op_8<op, uint8_t, uint8x16_t> },
         { "op_F32_F32_U8", &elementwise_comp_op_32<op, float, float32x4_t> },
         { "op_S16_S16_U8", &elementwise_comp_op_16<op, int16_t, int16x8_t> },
         { "op_S32_S32_U8", &elementwise_comp_op_32<op, int32_t, int32x4_t> },
+#endif /* defined(__ARM_FEATURE_SVE) */
         { "op_QASYMM8_SIGNED_QASYMM8_SIGNED_U8", &elementwise_comp_op_quantized_signed<op> },
         { "op_QASYMM8_QASYMM8_U8", &elementwise_comp_op_quantized<op> }
     };
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    map_function["op_F16_F16_U8"] = &elementwise_comp_op_16<op, float16_t, float16x8_t>;
+#if defined(__ARM_FEATURE_SVE)
+    map_function["op_F16_F16_U8"] = &arm_compute::cpu::sve::elementwise_comparison_op<op, float16_t>;
+#else  /* defined(__ARM_FEATURE_SVE) */
+    map_function["op_F16_F16_U8"]  = &elementwise_comp_op_16<op, float16_t, float16x8_t>;
+#endif /* defined(__ARM_FEATURE_SVE) */
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
     return configure_func(input1, input2, output, map_function);
diff --git a/src/core/NEON/wrapper/intrinsics/intrinsics.h b/src/core/NEON/wrapper/intrinsics/intrinsics.h
index 4c7b674e2e..871d9cc5ac 100644
--- a/src/core/NEON/wrapper/intrinsics/intrinsics.h
+++ b/src/core/NEON/wrapper/intrinsics/intrinsics.h
@@ -79,6 +79,7 @@
 #include "src/core/NEON/wrapper/intrinsics/svdup_n.h"
 #include "src/core/NEON/wrapper/intrinsics/svexp.h"
 #include "src/core/NEON/wrapper/intrinsics/svlog.h"
+#include "src/core/NEON/wrapper/intrinsics/svpow.h"
 #include "src/core/NEON/wrapper/intrinsics/svptrue.h"
 #include "src/core/NEON/wrapper/intrinsics/svqadd.h"
 #include "src/core/NEON/wrapper/intrinsics/svsin.h"
diff --git a/src/core/NEON/wrapper/svtraits.h b/src/core/NEON/wrapper/svtraits.h
new file mode 100644
index 0000000000..465983d16f
--- /dev/null
+++ b/src/core/NEON/wrapper/svtraits.h
@@ -0,0 +1,70 @@
+/*
+ * Copyright (c) 2021 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 SRC_CORE_NEON_WRAPPER_SVTRAITS_H
+#define SRC_CORE_NEON_WRAPPER_SVTRAITS_H
+#if defined(__ARM_FEATURE_SVE)
+#include "src/core/NEON/SVEMath.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace wrapper
+{
+template <typename T>
+struct sve_scalar;
+template <typename T>
+struct sve_vector;
+
+#define DEFINE_TYPES(stype)      \
+    template <>                  \
+    struct sve_scalar<sv##stype> \
+    {                            \
+        using type = stype;      \
+    };                           \
+    template <>                  \
+    struct sve_vector<stype>     \
+    {                            \
+        using type = sv##stype;  \
+    };
+
+DEFINE_TYPES(int8_t)
+DEFINE_TYPES(uint8_t)
+DEFINE_TYPES(int16_t)
+DEFINE_TYPES(uint16_t)
+DEFINE_TYPES(int32_t)
+DEFINE_TYPES(uint32_t)
+DEFINE_TYPES(int64_t)
+DEFINE_TYPES(uint64_t)
+DEFINE_TYPES(float16_t)
+DEFINE_TYPES(float32_t)
+DEFINE_TYPES(float64_t)
+DEFINE_TYPES(bfloat16_t)
+
+#undef DEFINE_TYPES
+
+} // namespace wrapper
+} // namespace arm_compute
+
+#endif /* defined(__ARM_FEATURE_SVE) */
+#endif /* #ifndef SRC_CORE_NEON_WRAPPER_SVTRAITS_H */
diff --git a/src/core/SVE/kernels/elementwise/impl/elementwise_list.h b/src/core/SVE/kernels/elementwise/impl/elementwise_list.h
new file mode 100644
index 0000000000..83c3355de4
--- /dev/null
+++ b/src/core/SVE/kernels/elementwise/impl/elementwise_list.h
@@ -0,0 +1,366 @@
+/*
+ * Copyright (c) 2021 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 SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H
+#define SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H
+#if defined(__ARM_FEATURE_SVE)
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/SVEMath.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "src/core/NEON/wrapper/svtraits.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace sve
+{
+using namespace arm_compute::wrapper;
+
+template <typename VectorType>
+inline VectorType elementwise_pow(svbool_t &pg, const VectorType &a, const VectorType &b)
+{
+    return svpow_z(pg, a, b);
+}
+
+template <>
+inline svint32_t elementwise_pow<svint32_t>(svbool_t &pg, const svint32_t &a, const svint32_t &b)
+{
+    return svcvt_s32_z(pg, svpow_z(pg, svcvt_f32_z(pg, a), svcvt_f32_z(pg, b)));
+}
+
+template <typename VectorType>
+inline VectorType elementwise_div(svbool_t &pg, const VectorType &a, const VectorType &b)
+{
+    return svdiv_z(pg, a, b);
+}
+
+template <>
+inline svint32_t elementwise_div<svint32_t>(svbool_t &pg, const svint32_t &a, const svint32_t &b)
+{
+    return svcvt_s32_z(pg, svdiv_z(pg, svcvt_f32_z(pg, a), svcvt_f32_z(pg, b)));
+}
+
+template <typename VectorType>
+inline VectorType elementwise_arithmetic_op(svbool_t &pg, const VectorType &a, const VectorType &b, ArithmeticOperation op)
+{
+    using ScalarType = typename sve_scalar<VectorType>::type;
+    VectorType res{};
+
+    switch(op)
+    {
+        case ArithmeticOperation::MAX:
+            res = svmax_z(pg, a, b);
+            break;
+        case ArithmeticOperation::MIN:
+            res = svmin_z(pg, a, b);
+            break;
+        case ArithmeticOperation::SQUARED_DIFF:
+        {
+            const auto tmp = svsub_z(pg, a, b);
+            res            = svmul_z(pg, tmp, tmp);
+            break;
+        }
+        case ArithmeticOperation::PRELU:
+        {
+            const auto zero = svdup_n(ScalarType(0));
+            const auto tmp  = svmul_z(pg, a, b);
+            const auto gt   = svcmpgt(pg, a, zero);
+            res             = svsel(gt, a, tmp);
+            break;
+        }
+        case ArithmeticOperation::DIV:
+        {
+            res = elementwise_div(pg, a, b);
+            break;
+        }
+        case ArithmeticOperation::POWER:
+        {
+            res = elementwise_pow(pg, a, b);
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+
+    return res;
+}
+
+template <uint32_t bytewidth>
+inline svbool_t narrow_to_byte_predicate(svbool_t pg)
+{
+    const auto all_false = svpfalse();
+
+    switch(bytewidth)
+    {
+        case 8:
+            pg = svuzp1_b32(pg, all_false);
+        /* fall through */
+        case 4:
+            pg = svuzp1_b16(pg, all_false);
+        /* fall through */
+        case 2:
+            pg = svuzp1_b8(pg, all_false);
+        /* fall through */
+        default:
+            break;
+    }
+    return pg;
+}
+
+template <typename InputVectorType, typename OutputVectorType>
+inline OutputVectorType elementwise_comparison_op(svbool_t &pg, const InputVectorType &a, const InputVectorType &b, ComparisonOperation op)
+{
+    svbool_t selection_vector{};
+
+    switch(op)
+    {
+        case ComparisonOperation::Equal:
+            selection_vector = svcmpeq(pg, a, b);
+            break;
+        case ComparisonOperation::NotEqual:
+            selection_vector = svcmpne(pg, a, b);
+            break;
+        case ComparisonOperation::Greater:
+            selection_vector = svcmpgt(pg, a, b);
+            break;
+        case ComparisonOperation::GreaterEqual:
+            selection_vector = svcmpge(pg, a, b);
+            break;
+        case ComparisonOperation::Less:
+            selection_vector = svcmplt(pg, a, b);
+            break;
+        case ComparisonOperation::LessEqual:
+            selection_vector = svcmple(pg, a, b);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+
+    using InputScalarType = typename sve_scalar<InputVectorType>::type;
+    selection_vector      = narrow_to_byte_predicate<sizeof(InputScalarType)>(selection_vector);
+
+    using OutputScalarType  = typename sve_scalar<OutputVectorType>::type;
+    const auto false_vector = svdup_n(static_cast<OutputScalarType>((uint32_t)0));
+    const auto true_vector  = svdup_n(static_cast<OutputScalarType>(~(uint32_t)0));
+    auto       ret          = svsel(selection_vector, true_vector, false_vector);
+
+    return ret;
+}
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+struct LoopArguments
+{
+    OperatorType           op;
+    const InputScalarType *input1_ptr;
+    const InputScalarType *input2_ptr;
+    OutputScalarType      *output_ptr;
+};
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+struct BroadcastLoopArguments
+{
+    OperatorType           op;
+    const InputScalarType *input1_ptr;
+    InputScalarType        broadcast_value;
+    OutputScalarType      *output_ptr;
+    bool                   reorder;
+};
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void arithmetic_op_loop(svbool_t pg, const LoopArguments<InputScalarType, OutputScalarType, ArithmeticOperation> &args)
+{
+    const auto in1 = svld1(pg, args.input1_ptr);
+    const auto in2 = svld1(pg, args.input2_ptr);
+    const auto res = elementwise_arithmetic_op<typename sve_vector<InputScalarType>::type>(pg, in1, in2, args.op);
+    svst1(pg, args.output_ptr, res);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void arithmetic_op_broadcast_loop(svbool_t pg, const BroadcastLoopArguments<InputScalarType, OutputScalarType, ArithmeticOperation> &args)
+{
+    const auto non_broadcast_vector = svld1(pg, args.input1_ptr);
+    const auto broadcast_vector     = svdup_n(args.broadcast_value);
+    const auto in1                  = args.reorder ? broadcast_vector : non_broadcast_vector;
+    const auto in2                  = args.reorder ? non_broadcast_vector : broadcast_vector;
+    const auto res                  = elementwise_arithmetic_op<typename sve_vector<InputScalarType>::type>(pg, in1, in2, args.op);
+    svst1(pg, args.output_ptr, res);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void comparison_op_loop(svbool_t pg, const LoopArguments<InputScalarType, OutputScalarType, ComparisonOperation> &args)
+{
+    const auto     in1       = svld1(pg, args.input1_ptr);
+    const auto     in2       = svld1(pg, args.input2_ptr);
+    const auto     res       = elementwise_comparison_op<typename sve_vector<InputScalarType>::type, typename sve_vector<OutputScalarType>::type>(pg, in1, in2, args.op);
+    const svbool_t output_pg = narrow_to_byte_predicate<sizeof(InputScalarType)>(pg);
+    svst1(output_pg, args.output_ptr, res);
+}
+
+template <typename InputScalarType, typename OutputScalarType>
+inline void comparison_op_broadcast_loop(svbool_t pg, const BroadcastLoopArguments<InputScalarType, OutputScalarType, ComparisonOperation> &args)
+{
+    const auto     non_broadcast_vector = svld1(pg, args.input1_ptr);
+    const auto     broadcast_vector     = svdup_n(args.broadcast_value);
+    const auto     in1                  = args.reorder ? broadcast_vector : non_broadcast_vector;
+    const auto     in2                  = args.reorder ? non_broadcast_vector : broadcast_vector;
+    const auto     res                  = elementwise_comparison_op<typename sve_vector<InputScalarType>::type, typename sve_vector<OutputScalarType>::type>(pg, in1, in2, args.op);
+    const svbool_t output_pg            = narrow_to_byte_predicate<sizeof(InputScalarType)>(pg);
+    svst1(output_pg, args.output_ptr, res);
+}
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+using LoopFuncType = void (*)(svbool_t, const LoopArguments<InputScalarType, OutputScalarType, OperatorType> &);
+
+template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
+using BroadcastLoopFuncType = void (*)(svbool_t, const BroadcastLoopArguments<InputScalarType, OutputScalarType, OperatorType> &);
+
+template <typename InputVectorType, typename OutputVectorType, typename OperatorType,
+          typename InputScalarType  = typename sve_scalar<InputVectorType>::type,
+          typename OutputScalarType = typename sve_scalar<OutputVectorType>::type>
+void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
+                    OperatorType op,
+                    LoopFuncType<InputScalarType, OutputScalarType, OperatorType>          func,
+                    BroadcastLoopFuncType<InputScalarType, OutputScalarType, OperatorType> broadcast_func)
+{
+    const auto all_true_pg = svptrue<InputScalarType>();
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            auto                  output_ptr              = reinterpret_cast<OutputScalarType *>(output.ptr());
+            const auto            non_broadcast_input_ptr = reinterpret_cast<const InputScalarType *>(non_broadcast_input.ptr());
+            const InputScalarType broadcast_value         = *reinterpret_cast<const InputScalarType *>(broadcast_input.ptr());
+
+            int x = window_start_x;
+
+            svbool_t pg = svwhilelt<InputScalarType>(x, window_end_x);
+            do
+            {
+                broadcast_func(pg,
+                {
+                    op,
+                    non_broadcast_input_ptr + x,
+                    broadcast_value,
+                    output_ptr + x,
+                    !is_broadcast_input_2
+                });
+                x += svcnt<InputScalarType>();
+                pg = svwhilelt<InputScalarType>(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            auto       output_ptr = reinterpret_cast<OutputScalarType *>(output.ptr());
+            const auto input1_ptr = reinterpret_cast<const InputScalarType *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const InputScalarType *>(input2.ptr());
+
+            int x = window_start_x;
+
+            svbool_t pg = svwhilelt<InputScalarType>(x, window_end_x);
+            do
+            {
+                func(pg,
+                {
+                    op,
+                    input1_ptr + x,
+                    input2_ptr + x,
+                    output_ptr + x
+                });
+                x += svcnt<InputScalarType>();
+                pg = svwhilelt<InputScalarType>(x, window_end_x);
+            }
+            while(svptest_any(all_true_pg, pg));
+        },
+        input1, input2, output);
+    }
+}
+
+template <ArithmeticOperation op, typename ScalarType>
+void elementwise_arithmetic_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    using VectorType = typename sve_vector<ScalarType>::type;
+
+    elementwise_op<VectorType, VectorType, ArithmeticOperation>(in1, in2, out, window, op,
+                                                                &arithmetic_op_loop<ScalarType, ScalarType>,
+                                                                &arithmetic_op_broadcast_loop<ScalarType, ScalarType>);
+}
+
+template <ComparisonOperation op, typename InputScalarType, typename OutputScalarType = uint8_t>
+void elementwise_comparison_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    static_assert(sizeof(InputScalarType) >= sizeof(OutputScalarType), "input data type's width should be equal to or greater than output data type's width");
+    using InputVectorType  = typename sve_vector<InputScalarType>::type;
+    using OutputVectorType = typename sve_vector<OutputScalarType>::type;
+
+    elementwise_op<InputVectorType, OutputVectorType, ComparisonOperation>(in1, in2, out, window, op,
+                                                                           &comparison_op_loop<InputScalarType, OutputScalarType>,
+                                                                           &comparison_op_broadcast_loop<InputScalarType, OutputScalarType>);
+}
+
+} // namespace sve
+} // namespace cpu
+} // namespace arm_compute
+#endif // defined(__ARM_FEATURE_SVE)
+#endif /* SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H */