Enable fat binary support

Changes our build system to allow building both Neon(TM) and SVE
kernels and package them in the same binary. This will allow
runtime selection of the underlying architecture.

Adds new build option, fat_binary, for enabling this feature.

Change-Id: I8e8386149773ce28e071a2fb7ddd8c8ae0f28a4a
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5704
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 466 insertions, 316 deletions

diff --git a/src/core/cpu/kernels/elementwise/sve/elementwise.cpp b/src/core/cpu/kernels/elementwise/sve/elementwise.cpp
new file mode 100644
index 0000000000..2c3bb0ff7c
--- /dev/null
+++ b/src/core/cpu/kernels/elementwise/sve/elementwise.cpp
@@ -0,0 +1,309 @@
+/*
+ * 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+#include "src/core/cpu/kernels/elementwise/sve/elementwise_list.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+using namespace arm_compute::wrapper;
+
+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>
+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>
+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>
+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>
+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>);
+}
+
+template <>
+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 <>
+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 <>
+svint16_t elementwise_div<svint16_t>(svbool_t &pg, const svint16_t &a, const svint16_t &b)
+{
+    ARM_COMPUTE_UNUSED(pg, a, b);
+    ARM_COMPUTE_ERROR("Not supported");
+}
+
+template void elementwise_arithmetic_op<ArithmeticOperation::MAX, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::MAX, float32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::MAX, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::MAX, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_arithmetic_op<ArithmeticOperation::MIN, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::MIN, float32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::MIN, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::MIN, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_arithmetic_op<ArithmeticOperation::SQUARED_DIFF, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::SQUARED_DIFF, float32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::SQUARED_DIFF, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::SQUARED_DIFF, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_arithmetic_op<ArithmeticOperation::PRELU, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::PRELU, float32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::PRELU, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::PRELU, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_arithmetic_op<ArithmeticOperation::DIV, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::DIV, float32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::DIV, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::DIV, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_arithmetic_op<ArithmeticOperation::POWER, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::POWER, float32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::POWER, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_arithmetic_op<ArithmeticOperation::POWER, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_comparison_op<ComparisonOperation::Equal, float>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Equal, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Equal, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Equal, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Equal, uint8_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_comparison_op<ComparisonOperation::NotEqual, float>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::NotEqual, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::NotEqual, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::NotEqual, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::NotEqual, uint8_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_comparison_op<ComparisonOperation::Greater, float>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Greater, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Greater, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Greater, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Greater, uint8_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_comparison_op<ComparisonOperation::GreaterEqual, float>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::GreaterEqual, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::GreaterEqual, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::GreaterEqual, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::GreaterEqual, uint8_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_comparison_op<ComparisonOperation::Less, float>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Less, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Less, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Less, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::Less, uint8_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+
+template void elementwise_comparison_op<ComparisonOperation::LessEqual, float>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::LessEqual, int32_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::LessEqual, float16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::LessEqual, int16_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+template void elementwise_comparison_op<ComparisonOperation::LessEqual, uint8_t>(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/elementwise/sve/elementwise_list.h b/src/core/cpu/kernels/elementwise/sve/elementwise_list.h
index 83c3355de4..a92a8648a8 100644
--- a/src/core/cpu/kernels/elementwise/sve/elementwise_list.h
+++ b/src/core/cpu/kernels/elementwise/sve/elementwise_list.h
@@ -23,50 +23,62 @@
  */
 #ifndef SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H
 #define SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H
-#if defined(__ARM_FEATURE_SVE)
+#if defined(ENABLE_SVE)
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/core/Window.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 "src/core/cpu/kernels/elementwise/sve/elementwise_list.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)
+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)
+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)
+template <uint32_t bytewidth>
+svbool_t narrow_to_byte_predicate(svbool_t pg)
 {
-    return svcvt_s32_z(pg, svdiv_z(pg, svcvt_f32_z(pg, a), svcvt_f32_z(pg, b)));
+    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 VectorType>
-inline VectorType elementwise_arithmetic_op(svbool_t &pg, const VectorType &a, const VectorType &b, ArithmeticOperation op)
+VectorType elementwise_arithmetic_op(svbool_t &pg, const VectorType &a, const VectorType &b, ArithmeticOperation op)
 {
-    using ScalarType = typename sve_scalar<VectorType>::type;
+    using ScalarType = typename wrapper::sve_scalar<VectorType>::type;
     VectorType res{};
 
     switch(op)
@@ -108,30 +120,8 @@ inline VectorType elementwise_arithmetic_op(svbool_t &pg, const VectorType &a, c
     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)
+OutputVectorType elementwise_comparison_op(svbool_t &pg, const InputVectorType &a, const InputVectorType &b, ComparisonOperation op)
 {
     svbool_t selection_vector{};
 
@@ -159,10 +149,10 @@ inline OutputVectorType elementwise_comparison_op(svbool_t &pg, const InputVecto
             ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
     }
 
-    using InputScalarType = typename sve_scalar<InputVectorType>::type;
+    using InputScalarType = typename wrapper::sve_scalar<InputVectorType>::type;
     selection_vector      = narrow_to_byte_predicate<sizeof(InputScalarType)>(selection_vector);
 
-    using OutputScalarType  = typename sve_scalar<OutputVectorType>::type;
+    using OutputScalarType  = typename wrapper::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);
@@ -170,197 +160,12 @@ inline OutputVectorType elementwise_comparison_op(svbool_t &pg, const InputVecto
     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>);
-}
+void elementwise_arithmetic_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
 
-} // namespace sve
+template <ComparisonOperation op, typename ScalarType, typename OutputScalarType = uint8_t>
+void elementwise_comparison_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window);
 } // namespace cpu
 } // namespace arm_compute
-#endif // defined(__ARM_FEATURE_SVE)
+#endif // defined(ENABLE_SVE)
 #endif /* SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H */
diff --git a/src/core/cpu/kernels/elementwise/sve/elementwise_quantized_list.h b/src/core/cpu/kernels/elementwise/sve/elementwise_quantized_list.h
index b6342c727c..6c5524e284 100644
--- a/src/core/cpu/kernels/elementwise/sve/elementwise_quantized_list.h
+++ b/src/core/cpu/kernels/elementwise/sve/elementwise_quantized_list.h
@@ -26,14 +26,13 @@
 
 #if defined(__ARM_FEATURE_SVE2)
 
+#include "src/core/NEON/wrapper/svtraits.h"
 #include "src/core/cpu/kernels/elementwise/sve/elementwise_list.h"
 
 namespace arm_compute
 {
 namespace cpu
 {
-namespace sve
-{
 using namespace arm_compute::wrapper;
 
 template <typename InputScalarType, typename OutputScalarType, typename OperatorType>
@@ -176,7 +175,7 @@ inline void comparison_op_quantized_loop(svbool_t pg, const QuantizedLoopArgumen
     const auto in1 = load_quantized(args.input1_ptr, pg, args.in1_offset, args.in1_scale);
     const auto in2 = load_quantized(args.input2_ptr, pg, args.in2_offset, args.in2_scale);
 
-    using OutputVectorType = typename sve_vector<OutputScalarType>::type;
+    using OutputVectorType = typename wrapper::traits::sve_vector<OutputScalarType>::type;
 
     const auto result = svcreate4(
                             elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 0), svget4(in2, 0), args.op),
@@ -200,7 +199,7 @@ inline void comparison_op_broadcast_quantized_loop(svbool_t pg, const BroadcastQ
     const auto &af = args.reorder ? in2 : in1;
     const auto &bf = args.reorder ? in1 : in2;
 
-    using OutputVectorType = typename sve_vector<OutputScalarType>::type;
+    using OutputVectorType = typename wrapper::traits::sve_vector<OutputScalarType>::type;
 
     const auto result = svcreate4(
                             elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(af, 0), svget4(bf, 0), args.op),
@@ -221,8 +220,8 @@ template <typename InputScalarType, typename OutputScalarType, typename Operator
 using BroadcastQuantizedLoopFuncType = void (*)(svbool_t, const BroadcastQuantizedLoopArguments<InputScalarType, OutputScalarType, OperatorType> &);
 
 template <typename InputVectorType, typename OutputVectorType, typename OperatorType,
-          typename InputScalarType  = typename sve_scalar<InputVectorType>::type,
-          typename OutputScalarType = typename sve_scalar<OutputVectorType>::type>
+          typename InputScalarType  = typename wrapper::sve_scalar<InputVectorType>::type,
+          typename OutputScalarType = typename wrapper::sve_scalar<OutputVectorType>::type>
 void elementwise_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
                               OperatorType op,
                               LoopQuantizedFuncType<InputScalarType, OutputScalarType, OperatorType>          func,
@@ -344,7 +343,7 @@ void elementwise_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *o
 template <ArithmeticOperation op, typename ScalarType>
 void elementwise_arithmetic_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    using VectorType = typename sve_vector<ScalarType>::type;
+    using VectorType = typename wrapper::traits::sve_vector<ScalarType>::type;
     elementwise_quantized_op<VectorType, VectorType, ArithmeticOperation>(in1, in2, out, window, op,
                                                                           &arithmetic_op_quantized_loop<ScalarType, ScalarType>,
                                                                           &arithmetic_op_broadcast_quantized_loop<ScalarType, ScalarType>);
@@ -354,14 +353,12 @@ template <ComparisonOperation op, typename InputScalarType, typename OutputScala
 void elementwise_comparison_quantized_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;
+    using InputVectorType  = typename wrapper::traits::sve_vector<InputScalarType>::type;
+    using OutputVectorType = typename wrapper::traits::sve_vector<OutputScalarType>::type;
     elementwise_quantized_op<InputVectorType, OutputVectorType, ComparisonOperation>(in1, in2, out, window, op,
                                                                                      &comparison_op_quantized_loop<InputScalarType, OutputScalarType>,
                                                                                      &comparison_op_broadcast_quantized_loop<InputScalarType, OutputScalarType>);
 }
-
-} // namespace sve
 } // namespace cpu
 } // namespace arm_compute
 
diff --git a/src/core/cpu/kernels/elementwise/sve/elementwise_unary.cpp b/src/core/cpu/kernels/elementwise/sve/elementwise_unary.cpp
new file mode 100644
index 0000000000..cb58548f0b
--- /dev/null
+++ b/src/core/cpu/kernels/elementwise/sve/elementwise_unary.cpp
@@ -0,0 +1,111 @@
+/*
+ * 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.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/SVEMath.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <typename ScalarType, typename VectorType>
+inline typename std::enable_if<utils::traits::is_floating_point<ScalarType>::value, VectorType>::type elementwise_op_sve_imp(svbool_t pg, ElementWiseUnary op, const VectorType &a)
+{
+    switch(op)
+    {
+        case ElementWiseUnary::RSQRT:
+            return svinvsqrt(pg, a);
+        case ElementWiseUnary::EXP:
+            return wrapper::svexp_z(pg, a);
+        case ElementWiseUnary::NEG:
+            return svneg_z(pg, a);
+        case ElementWiseUnary::LOG:
+            return wrapper::svlog_z(pg, a);
+        case ElementWiseUnary::ABS:
+            return svabs_z(pg, a);
+        case ElementWiseUnary::ROUND:
+            return svrintn_z(pg, a);
+        case ElementWiseUnary::SIN:
+            return wrapper::svsin_z(pg, a);
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED");
+    }
+}
+
+template <typename ScalarType, typename VectorType>
+inline typename std::enable_if<std::is_integral<ScalarType>::value, VectorType>::type elementwise_op_sve_imp(svbool_t pg, ElementWiseUnary op, const VectorType &a)
+{
+    switch(op)
+    {
+        case ElementWiseUnary::NEG:
+            return svneg_z(pg, a);
+        case ElementWiseUnary::ABS:
+            return svabs_z(pg, a);
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED");
+    }
+}
+
+template <typename ScalarType>
+void elementwise_sve_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
+{
+    const auto all_true_pg    = wrapper::svptrue<ScalarType>();
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(in, win);
+    Iterator output(out, win);
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        auto       output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
+        const auto input_ptr  = reinterpret_cast<const ScalarType *>(input.ptr());
+        int        x          = window_start_x;
+
+        svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
+        do
+        {
+            const auto vin = svld1(pg, input_ptr + x);
+            svst1(pg, output_ptr + x, elementwise_op_sve_imp<ScalarType, decltype(vin)>(pg, op, vin));
+            x += wrapper::svcnt<ScalarType>();
+            pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
+        }
+        while(svptest_any(all_true_pg, pg));
+    },
+    input, output);
+}
+
+template void elementwise_sve_op<float16_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
+template void elementwise_sve_op<float32_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
+template void elementwise_sve_op<int32_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/elementwise/sve/elementwise_unary_list.h b/src/core/cpu/kernels/elementwise/sve/elementwise_unary_list.h
index 23502c71e5..63490421e9 100644
--- a/src/core/cpu/kernels/elementwise/sve/elementwise_unary_list.h
+++ b/src/core/cpu/kernels/elementwise/sve/elementwise_unary_list.h
@@ -25,87 +25,15 @@
 #define SRC_CORE_SVE_KERNELS_ELEMENTWISE_UNARY_LIST_H
 
 #include "arm_compute/core/Types.h"
-#include "arm_compute/core/utils/misc/Traits.h"
-#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
-#if defined(__ARM_FEATURE_SVE)
-#include "src/core/NEON/SVEMath.h"
-#include <arm_sve.h>
+#if defined(ENABLE_SVE)
 
 namespace arm_compute
 {
 namespace cpu
 {
-template <typename ScalarType, typename VectorType>
-inline typename std::enable_if<utils::traits::is_floating_point<ScalarType>::value, VectorType>::type elementwise_op_sve_imp(svbool_t pg, ElementWiseUnary op, const VectorType &a)
-{
-    switch(op)
-    {
-        case ElementWiseUnary::RSQRT:
-            return svinvsqrt(pg, a);
-        case ElementWiseUnary::EXP:
-            return wrapper::svexp_z(pg, a);
-        case ElementWiseUnary::NEG:
-            return svneg_z(pg, a);
-        case ElementWiseUnary::LOG:
-            return wrapper::svlog_z(pg, a);
-        case ElementWiseUnary::ABS:
-            return svabs_z(pg, a);
-        case ElementWiseUnary::ROUND:
-            return svrintn_z(pg, a);
-        case ElementWiseUnary::SIN:
-            return wrapper::svsin_z(pg, a);
-        default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED");
-    }
-}
-
-template <typename ScalarType, typename VectorType>
-inline typename std::enable_if<std::is_integral<ScalarType>::value, VectorType>::type elementwise_op_sve_imp(svbool_t pg, ElementWiseUnary op, const VectorType &a)
-{
-    switch(op)
-    {
-        case ElementWiseUnary::NEG:
-            return svneg_z(pg, a);
-        case ElementWiseUnary::ABS:
-            return svabs_z(pg, a);
-        default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED");
-    }
-}
-
 template <typename ScalarType>
-void elementwise_sve_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
-{
-    const auto all_true_pg    = wrapper::svptrue<ScalarType>();
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
-
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    Iterator input(in, win);
-    Iterator output(out, win);
-
-    execute_window_loop(win, [&](const Coordinates &)
-    {
-        auto       output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
-        const auto input_ptr  = reinterpret_cast<const ScalarType *>(input.ptr());
-        int        x          = window_start_x;
-
-        svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
-        do
-        {
-            const auto vin = svld1(pg, input_ptr + x);
-            svst1(pg, output_ptr + x, elementwise_op_sve_imp<ScalarType, decltype(vin)>(pg, op, vin));
-            x += wrapper::svcnt<ScalarType>();
-            pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
-        }
-        while(svptest_any(all_true_pg, pg));
-    },
-    input, output);
-}
-
+void elementwise_sve_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
 } // namespace cpu
 } // namespace arm_compute
-#endif // defined(__ARM_FEATURE_SVE)
+#endif // defined(ENABLE_SVE)
 #endif // SRC_CORE_NEON_KERNELS_ELEMENTWISE_UNARY_LIST_H
\ No newline at end of file