COMPMID-3874: Create ArithmeticAddition SVE/SVE2

Change-Id: I4ec7561a7f6a42a22b8187968ae302dbe75023bc
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4753
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 137 insertions, 806 deletions

diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
index aa7af54e9c..f706ee5694 100644
--- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -28,7 +28,11 @@
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/Validate.h"
 #include "src/core/CPP/Validate.h"
+#include "src/core/NEON/kernels/arithmetic_addition/impl/NEON/list.h"
+#include "src/core/NEON/kernels/arithmetic_addition/impl/SVE/list.h"
 #include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Registrars.h"
+#include "src/core/common/StdTypes.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
 
@@ -39,788 +43,156 @@ namespace arm_compute
 {
 namespace
 {
-template <typename T>
-void add_same(const ITensor *in1, const ITensor *in2, ITensor *out, const ConvertPolicy policy, const Window &window)
+struct ArithmeticAdditionSelectorData
 {
-    /** NEON vector tag type. */
-    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
+    DataType dt1;
+    DataType dt2;
+    DataType dt3;
+};
 
-    // 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());
+using ArithmeticAdditionSelectorPtr = std::add_pointer<bool(const ArithmeticAdditionSelectorData &data)>::type;
 
-    // Clear X Dimension on execution window as we handle manually
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    constexpr int window_step_x         = 16 / sizeof(T);
-    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();
+struct ArithmeticAdditionKernel
+{
+    const char                                             *name;
+    const ArithmeticAdditionSelectorPtr                     is_selected;
+    NEArithmeticAdditionKernel::ArithmeticAdditionKernelPtr ukernel;
+};
 
-    if(is_broadcast_across_x)
+static const ArithmeticAdditionKernel available_kernels[] =
+{
+#if defined(__ARM_FEATURE_SVE)
     {
-        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 &)
-        {
-            const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr());
-            const auto output_ptr              = reinterpret_cast<T *>(output.ptr());
-
-            const T    broadcast_value     = *reinterpret_cast<const T *>(broadcast_input.ptr());
-            const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
-                const auto res             = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(broadcast_value_vec, non_broadcast_v) : wrapper::vadd(broadcast_value_vec, non_broadcast_v);
-                wrapper::vstore(output_ptr + x, res);
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
-                *(output_ptr + x)          = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(broadcast_value, non_broadcast_v) : broadcast_value + non_broadcast_v;
-            }
-        },
-        broadcast_input, non_broadcast_input, output);
-    }
-    else
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F32)); },
+        REGISTER_FP32_SVE(arm_compute::cpu::arithmetic_addition_same_sve<float>)
+    },
     {
-        // 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 &)
-        {
-            const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr());
-            const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr());
-            const auto output_ptr = reinterpret_cast<T *>(output.ptr());
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto val1 = wrapper::vloadq(input1_ptr + x);
-                const auto val2 = wrapper::vloadq(input2_ptr + x);
-                const auto res  = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(val1, val2) : wrapper::vadd(val1, val2);
-                wrapper::vstore(output_ptr + x, res);
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const auto val1   = *(input1_ptr + x);
-                const auto val2   = *(input2_ptr + x);
-                *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(val1, val2) : val1 + val2;
-            }
-        },
-        input1, input2, output);
-    }
-}
-
-void add_QASYMM8_QASYMM8_QASYMM8(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
-{
-    ARM_COMPUTE_UNUSED(policy);
-
-    // 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 int  window_step_x         = 16;
-    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();
-
-    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
-    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
-    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
-
-    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
-    const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
-
-    if(is_broadcast_across_x)
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F16)); },
+        REGISTER_FP16_SVE(arm_compute::cpu::arithmetic_addition_same_sve<float16_t>)
+    },
     {
-        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;
-        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
-        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
-
-        const float32x4_t vscale1  = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
-        const float32x4_t vscale2  = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
-        const int32x4_t   voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
-        const int32x4_t   voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
-
-        // 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 &)
-        {
-            const auto non_broadcast_input_ptr = reinterpret_cast<const uint8_t *>(non_broadcast_input.ptr());
-            const auto output_ptr              = reinterpret_cast<uint8_t *>(output.ptr());
-
-            const uint8_t    broadcast_value     = *reinterpret_cast<const uint8_t *>(broadcast_input.ptr());
-            const uint8x16_t broadcast_value_vec = vdupq_n_u8(broadcast_value);
-
-            const float32x4x4_t bf =
-            {
-                {
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2),
-                }
-            };
-            const float bfs = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const uint8x16_t    a = vld1q_u8(non_broadcast_input_ptr + x);
-                const float32x4x4_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
-                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
-                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
-                *(output_ptr + x) = quantize_qasymm8((afs + bfs), oq_info);
-            }
-        },
-        broadcast_input, non_broadcast_input, output);
-    }
-    else
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::U8)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_same_sve<uint8_t>)
+    },
     {
-        // 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);
-
-        const float32x4_t vscale1  = vdupq_n_f32(iq1_info.scale);
-        const float32x4_t vscale2  = vdupq_n_f32(iq2_info.scale);
-        const int32x4_t   voffset1 = vdupq_n_s32(iq1_info.offset);
-        const int32x4_t   voffset2 = vdupq_n_s32(iq2_info.offset);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
-            const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
-            const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const uint8x16_t a = vld1q_u8(input1_ptr + x);
-                const uint8x16_t b = vld1q_u8(input2_ptr + x);
-
-                const float32x4x4_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1),
-                    }
-                };
-
-                const float32x4x4_t bf =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
-                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
-                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale;
-                const float bfs   = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale;
-                *(output_ptr + x) = quantize_qasymm8((afs + bfs), out->info()->quantization_info());
-            }
-        },
-        input1, input2, output);
-    }
-}
-
-void add_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
-{
-    ARM_COMPUTE_UNUSED(policy);
-
-    // 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 int  window_step_x         = 16;
-    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();
-
-    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
-    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
-    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
-
-    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
-    const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
-
-    if(is_broadcast_across_x)
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S16)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_same_sve<int16_t>)
+    },
     {
-        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;
-        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
-        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
-
-        const float32x4_t vscale1  = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
-        const float32x4_t vscale2  = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
-        const int32x4_t   voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
-        const int32x4_t   voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
-
-        // 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 &)
-        {
-            const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
-            const auto output_ptr              = reinterpret_cast<int8_t *>(output.ptr());
-
-            const int8_t    broadcast_value     = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
-            const int8x16_t broadcast_value_vec = vdupq_n_s8(broadcast_value);
-
-            const float32x4x4_t bf =
-            {
-                {
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(broadcast_value_vec)))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(broadcast_value_vec)))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(broadcast_value_vec)))), voffset2)), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(broadcast_value_vec)))), voffset2)), vscale2),
-                }
-            };
-            const float bfs = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const int8x16_t     a = vld1q_s8(non_broadcast_input_ptr + x);
-                const float32x4x4_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
-                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
-                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
-                *(output_ptr + x) = quantize_qasymm8_signed((afs + bfs), oq_info);
-            }
-        },
-        broadcast_input, non_broadcast_input, output);
-    }
-    else
+        "arithmetic_addition_same_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S32)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_same_sve<int32_t>)
+    },
     {
-        // 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);
-
-        const float32x4_t vscale1  = vdupq_n_f32(iq1_info.scale);
-        const float32x4_t vscale2  = vdupq_n_f32(iq2_info.scale);
-        const int32x4_t   voffset1 = vdupq_n_s32(iq1_info.offset);
-        const int32x4_t   voffset2 = vdupq_n_s32(iq2_info.offset);
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr());
-            const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr());
-            const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const int8x16_t a = vld1q_s8(input1_ptr + x);
-                const int8x16_t b = vld1q_s8(input2_ptr + x);
-
-                const float32x4x4_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(a)))), voffset1)), vscale1),
-                    }
-                };
-
-                const float32x4x4_t bf =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(b)))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(b)))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(b)))), voffset2)), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(b)))), voffset2)), vscale2),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
-                        vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
-                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
-                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale;
-                const float bfs   = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale;
-                *(output_ptr + x) = quantize_qasymm8_signed((afs + bfs), out->info()->quantization_info());
-            }
-        },
-        input1, input2, output);
-    }
-}
-
-void add_QSYMM16_QSYMM16_QSYMM16(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
-{
-    ARM_COMPUTE_UNUSED(policy);
-
-    // 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 int  window_step_x         = 8;
-    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();
-
-    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
-    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
-    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
-
-    const float32x4_t vscale1    = vdupq_n_f32(iq1_info.scale);
-    const float32x4_t vscale2    = vdupq_n_f32(iq2_info.scale);
-    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
-
-    if(is_broadcast_across_x)
+        "arithmetic_addition_U8_S16_S16_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == DataType::U8) && (data.dt2 == DataType::S16)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_U8_S16_S16_sve)
+    },
     {
-        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;
-        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
-        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
-
-        // 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 &)
-        {
-            const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr());
-            const auto output_ptr              = reinterpret_cast<int16_t *>(output.ptr());
-
-            const int16_t   broadcast_value     = *reinterpret_cast<const int16_t *>(broadcast_input.ptr());
-            const int16x8_t broadcast_value_vec = vdupq_n_s16(broadcast_value);
-
-            const float32x4x2_t bf =
-            {
-                {
-                    vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(broadcast_value_vec))), vscale2),
-                    vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(broadcast_value_vec))), vscale2),
-                }
-            };
-            const float bfs = static_cast<int32_t>(broadcast_value) * broadcast_qinfo.scale;
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const int16x8_t     a = vld1q_s16(non_broadcast_input_ptr + x);
-                const float32x4x2_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1),
-                    }
-                };
-
-                const int32x4x4_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1]));
-                vst1q_s16(output_ptr + x, pa);
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x)) * non_broadcast_qinfo.scale;
-                *(output_ptr + x) = quantize_qsymm16((afs + bfs), oq_info);
-            }
-        },
-        broadcast_input, non_broadcast_input, output);
-    }
-    else
+        "arithmetic_addition_S16_U8_S16_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == DataType::S16) && (data.dt2 == DataType::U8)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_S16_U8_S16_sve)
+    },
     {
-        // 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 &)
-        {
-            const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
-            const auto input2_ptr = reinterpret_cast<const int16_t *>(input2.ptr());
-            const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const int16x8_t a = vld1q_s16(input1_ptr + x);
-                const int16x8_t b = vld1q_s16(input2_ptr + x);
-
-                const float32x4x2_t af =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1),
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1),
-                    }
-                };
-
-                const float32x4x2_t bf =
-                {
-                    {
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(b))), vscale2),
-                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(b))), vscale2),
-                    }
-                };
-
-                const int32x4x2_t rf =
-                {
-                    {
-#ifdef __aarch64__
-                        vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtnq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-#else  //__aarch64__
-                        vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
-                        vcvtq_s32_f32(vmulq_f32(vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
-#endif //__aarch64__
-                    }
-                };
-
-                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1]));
-                vst1q_s16(output_ptr + x, pa);
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const float afs   = static_cast<int32_t>((*(input1_ptr + x))) * iq1_info.scale;
-                const float bfs   = static_cast<int32_t>((*(input2_ptr + x))) * iq2_info.scale;
-                *(output_ptr + x) = quantize_qsymm16((afs + bfs), out->info()->quantization_info());
-            }
-        },
-        input1, input2, output);
-    }
-}
-
-void add_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
-{
-    // Create input windows
-    Window win        = window;
-    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
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    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);
-
-    const int  window_step_x  = 8;
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
-
-    execute_window_loop(win, [&](const Coordinates &)
+        "arithmetic_addition_U8_U8_S16_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt3 == DataType::S16)); },
+        REGISTER_INTEGER_SVE(arm_compute::cpu::arithmetic_addition_U8_U8_S16_sve)
+    },
+#else /* !defined(__ARM_FEATURE_SVE) */
     {
-        const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
-        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
-        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
-
-        if(policy == ConvertPolicy::WRAP)
-        {
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto vin1 = wrapper::vloadq(input1_ptr + x);
-                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
-                wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                *(output_ptr + x) = *(input1_ptr + x) + static_cast<int16_t>(*(input2_ptr + x));
-            }
-        }
-        else
-        {
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto vin1 = wrapper::vloadq(input1_ptr + x);
-                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
-                wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2));
-            }
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F32)); },
+        REGISTER_FP32_NEON(arm_compute::cpu::arithmetic_addition_same_neon<float>)
+    },
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    {
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F16)); },
+        REGISTER_FP16_NEON(arm_compute::cpu::arithmetic_addition_same_neon<float16_t>)
+    },
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
+    {
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::U8)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_same_neon<uint8_t>)
+    },
+    {
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_same_neon<int16_t>)
+    },
+    {
+        "arithmetic_addition_same_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S32)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_same_neon<int32_t>)
+    },
+    {
+        "arithmetic_addition_U8_S16_S16_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == DataType::U8) && (data.dt2 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_U8_S16_S16_neon)
+    },
+    {
+        "arithmetic_addition_S16_U8_S16_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == DataType::S16) && (data.dt2 == DataType::U8)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_S16_U8_S16_neon)
+    },
+    {
+        "arithmetic_addition_U8_U8_S16_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt3 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::arithmetic_addition_U8_U8_S16_neon)
+    },
+#endif /* defined(__ARM_FEATURE_SVE) */
 
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                *(output_ptr + x) = wrapper::add_sat(*(input1_ptr + x), static_cast<int16_t>(*(input2_ptr + x)));
-            }
-        }
+#if defined(__ARM_FEATURE_SVE2)
+    {
+        "arithmetic_addition_qasymm8_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8)); },
+        REGISTER_QASYMM8_SVE(arm_compute::cpu::arithmetic_addition_qasymm8_sve)
     },
-    input1, input2, output);
-}
+    {
+        "arithmetic_addition_qasymm8_signed_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8_SIGNED)); },
+        REGISTER_QASYMM8_SIGNED_SVE(arm_compute::cpu::arithmetic_addition_qasymm8_signed_sve)
+    },
+    {
+        "arithmetic_addition_qsymm16_sve",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QSYMM16)); },
+        REGISTER_QSYMM16_SVE(arm_compute::cpu::arithmetic_addition_qsymm16_sve)
+    },
+#else  /* !defined(__ARM_FEATURE_SVE2) */
+    {
+        "arithmetic_addition_qasymm8_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8)); },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::arithmetic_addition_qasymm8_neon)
+    },
+    {
+        "arithmetic_addition_qasymm8_signed_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8_SIGNED)); },
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::arithmetic_addition_qasymm8_signed_neon)
+    },
+    {
+        "arithmetic_addition_qsymm16_neon",
+        [](const ArithmeticAdditionSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QSYMM16)); },
+        REGISTER_QSYMM16_NEON(arm_compute::cpu::arithmetic_addition_qsymm16_neon)
+    },
+#endif /* defined(__ARM_FEATURE_SVE2) */
 
-inline void add_U8_S16_S16(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy, const Window &window)
-{
-    // Simply swap the two input buffers:
-    add_S16_U8_S16(input2, input1, output, policy, window);
-}
+};
 
-void add_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, ConvertPolicy policy, const Window &window)
+const ArithmeticAdditionKernel *get_implementation(DataType dt1, DataType dt2, DataType dt3)
 {
-    // Create input windows
-    Window win        = window;
-    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
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    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);
-
-    const int  window_step_x  = 8;
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
-
-    execute_window_loop(win, [&](const Coordinates &)
+    for(const auto &uk : available_kernels)
     {
-        const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
-        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
-        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
-
-        if(policy == ConvertPolicy::WRAP)
-        {
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
-                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
-                wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                *(output_ptr + x) = static_cast<int16_t>(*(input1_ptr + x)) + static_cast<int16_t>(*(input2_ptr + x));
-            }
-        }
-        else
+        if(uk.is_selected({ dt1, dt2, dt3 }))
         {
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
-                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
-                wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                *(output_ptr + x) = wrapper::add_sat(static_cast<int16_t>(*(input1_ptr + x)),
-                                                     static_cast<int16_t>(*(input2_ptr + x)));
-            }
+            return &uk;
         }
-    },
-    input1, input2, output);
+    }
+    return nullptr;
 }
 
 Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output, ConvertPolicy policy)
@@ -926,53 +298,12 @@ void NEArithmeticAdditionKernel::configure(const ITensorInfo *input1, const ITen
     ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1, *input2, *output, policy));
 
+    _policy = policy;
+    _func   = get_implementation(input1->data_type(), input2->data_type(), output->data_type())->ukernel;
+
     // Configure kernel window
     auto win_config = validate_and_configure_window(*input1, *input2, *output);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-
-    static std::map<std::string, AddFunction *> map_function =
-    {
-        { "add_wrap_QASYMM8_QASYMM8_QASYMM8", &add_QASYMM8_QASYMM8_QASYMM8 },
-        { "add_saturate_QASYMM8_QASYMM8_QASYMM8", &add_QASYMM8_QASYMM8_QASYMM8 },
-        { "add_wrap_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &add_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED },
-        { "add_saturate_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &add_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED },
-        { "add_wrap_QSYMM16_QSYMM16_QSYMM16", &add_QSYMM16_QSYMM16_QSYMM16 },
-        { "add_saturate_QSYMM16_QSYMM16_QSYMM16", &add_QSYMM16_QSYMM16_QSYMM16 },
-        { "add_wrap_U8_U8_U8", &add_same<uint8_t> },
-        { "add_saturate_U8_U8_U8", &add_same<uint8_t> },
-        { "add_wrap_S16_U8_S16", &add_S16_U8_S16 },
-        { "add_saturate_S16_U8_S16", &add_S16_U8_S16 },
-        { "add_wrap_U8_S16_S16", &add_U8_S16_S16 },
-        { "add_saturate_U8_S16_S16", &add_U8_S16_S16 },
-        { "add_wrap_U8_U8_S16", &add_U8_U8_S16 },
-        { "add_saturate_U8_U8_S16", &add_U8_U8_S16 },
-        { "add_wrap_S16_S16_S16", &add_same<int16_t> },
-        { "add_saturate_S16_S16_S16", &add_same<int16_t> },
-        { "add_wrap_S32_S32_S32", &add_same<int32_t> },
-        { "add_saturate_S32_S32_S32", &add_same<int32_t> },
-        { "add_wrap_F32_F32_F32", &add_same<float> },
-        { "add_saturate_F32_F32_F32", &add_same<float> },
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        { "add_wrap_F16_F16_F16", &add_same<float16_t> },
-        { "add_saturate_F16_F16_F16", &add_same<float16_t> },
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-    };
-
-    _policy = policy;
-
-    std::string function_to_call("add_");
-    function_to_call += policy == ConvertPolicy::WRAP ? "wrap_" : "saturate_";
-    function_to_call += string_from_data_type(input1->data_type()) + "_";
-    function_to_call += string_from_data_type(input2->data_type()) + "_";
-    function_to_call += string_from_data_type(output->data_type());
-
-    auto it = map_function.find(function_to_call);
-
-    if(it != map_function.end())
-    {
-        _func = it->second;
-    }
-
     INEKernel::configure(win_config.second);
 }