Multi ISA Technical Debt

 * Update json struct meet multi-ISA updates
 * Add impl.cpp in kernels where we only have impl.h

Resolves  COMPMID-5173
Change-Id: I5da3c4b016a5d0115c4ba46cbfefde7bce518ac1
Signed-off-by: Dana Zlotnik <dana.zlotnik@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7191
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

17 files changed, 771 insertions, 642 deletions

diff --git a/Android.bp b/Android.bp
index 3a49b8c362..a279fdf5bb 100644
--- a/Android.bp
+++ b/Android.bp
@@ -458,6 +458,7 @@ cc_library_static {
         "src/cpu/kernels/elementwise_binary/generic/neon/qasymm8_signed.cpp",
         "src/cpu/kernels/elementwise_unary/generic/neon/fp16.cpp",
         "src/cpu/kernels/elementwise_unary/generic/neon/fp32.cpp",
+        "src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp",
         "src/cpu/kernels/elementwise_unary/generic/neon/integer.cpp",
         "src/cpu/kernels/floor/neon/fp16.cpp",
         "src/cpu/kernels/floor/neon/fp32.cpp",
@@ -511,6 +512,7 @@ cc_library_static {
         "src/cpu/kernels/select/generic/neon/integer.cpp",
         "src/cpu/kernels/softmax/generic/neon/fp16.cpp",
         "src/cpu/kernels/softmax/generic/neon/fp32.cpp",
+        "src/cpu/kernels/softmax/generic/neon/impl.cpp",
         "src/cpu/kernels/softmax/generic/neon/qasymm8.cpp",
         "src/cpu/kernels/softmax/generic/neon/qasymm8_signed.cpp",
         "src/cpu/kernels/sub/neon/qasymm8.cpp",
diff --git a/filelist.json b/filelist.json
index 7e47df959c..3bdc00aeef 100644
--- a/filelist.json
+++ b/filelist.json
@@ -837,14 +837,14 @@
           "common": [
             "src/cpu/operators/CpuActivation.cpp",
             "src/cpu/kernels/CpuActivationKernel.cpp",
-            "src/runtime/NEON/functions/NEActivationLayer.cpp",
-            "src/cpu/kernels/activation/generic/neon/qasymm8.cpp",
-            "src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp",
-            "src/cpu/kernels/activation/generic/neon/qsymm16.cpp"
+            "src/runtime/NEON/functions/NEActivationLayer.cpp"
           ],
           "neon": {
             "fp16": [ "src/cpu/kernels/activation/generic/neon/fp16.cpp" ],
-            "fp32": [ "src/cpu/kernels/activation/generic/neon/fp32.cpp" ]
+            "fp32": [ "src/cpu/kernels/activation/generic/neon/fp32.cpp" ], 
+            "qasymm8": [ "src/cpu/kernels/activation/generic/neon/qasymm8.cpp" ],
+            "qasymm8_signed": [ "src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp" ],
+            "qsymm16": [ "src/cpu/kernels/activation/generic/neon/qsymm16.cpp" ]
           },
           "sve": {
             "fp16": [ "src/cpu/kernels/activation/generic/sve/fp16.cpp" ],
@@ -961,7 +961,6 @@
             "fp32":["src/cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp"],
             "fp16":["src/cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp"],
             "qsymm16":["src/cpu/kernels/boundingboxtransform/generic/neon/qsymm16.cpp"]
-
           }
         }
       },
@@ -1304,6 +1303,7 @@
             "src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp"
           ],
           "neon": {
+            "common":["src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp"],
             "integer": ["src/cpu/kernels/elementwise_unary/generic/neon/integer.cpp"],
             "fp32": ["src/cpu/kernels/elementwise_unary/generic/neon/fp32.cpp"],
             "fp16": ["src/cpu/kernels/elementwise_unary/generic/neon/fp16.cpp"]
@@ -1709,8 +1709,6 @@
             "src/cpu/operators/CpuPool2d.cpp",
             "src/cpu/kernels/CpuPool2dKernel.cpp",
             "src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp",
-            "src/cpu/kernels/pool2d/neon/qasymm8.cpp",
-            "src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp",
             "src/runtime/NEON/functions/NEPoolingLayer.cpp"
           ],
           "neon": {
@@ -1743,7 +1741,9 @@
             ],
             "nchw": [ "src/cpu/kernels/pool2d/neon/nchw/all.cpp" ],
             "fp16": [ "src/cpu/kernels/pool2d/neon/fp16.cpp" ],
-            "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ]
+            "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ], 
+            "qasymm8":[ "src/cpu/kernels/pool2d/neon/qasymm8.cpp" ], 
+            "qasymm8_signed":["src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp"]
           },
           "sve": {
             "common": [
@@ -1884,7 +1884,6 @@
             "integer": [ "src/cpu/kernels/scale/sve/integer.cpp" ],
             "qasymm8": [ "src/cpu/kernels/scale/sve/qasymm8.cpp" ],
             "qasymm8_signed": [ "src/cpu/kernels/scale/sve/qasymm8_signed.cpp" ]
-
           },
           "neon": {
             "fp16": [ "src/cpu/kernels/scale/neon/fp16.cpp" ],
@@ -1925,6 +1924,7 @@
             "src/runtime/NEON/functions/NESoftmaxLayer.cpp"
           ],
           "neon":{
+            "common":["src/cpu/kernels/softmax/generic/neon/impl.cpp"],
             "fp32": ["src/cpu/kernels/softmax/generic/neon/fp32.cpp"],
             "fp16": ["src/cpu/kernels/softmax/generic/neon/fp16.cpp"], 
             "qasymm8":[ "src/cpu/kernels/softmax/generic/neon/qasymm8.cpp"], 
@@ -1938,6 +1938,7 @@
             "qasymm8_signed": ["src/cpu/kernels/softmax/generic/sve/qasymm8_signed.cpp"]
           },
           "sve2":{
+            "common" :["src/cpu/kernels/softmax/generic/sve2/impl.cpp"],
             "qasymm8":[ "src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp"], 
             "qasymm8_signed":["src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp"]
           }
diff --git a/src/core/NEON/kernels/NECropKernel.cpp b/src/core/NEON/kernels/NECropKernel.cpp
index 729402116d..94c455305c 100644
--- a/src/core/NEON/kernels/NECropKernel.cpp
+++ b/src/core/NEON/kernels/NECropKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,7 +35,7 @@
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
 #include "src/core/utils/helpers/bit_ops.h"
-#include "src/cpu/kernels/crop/generic/neon/list.h"
+#include "src/cpu/kernels/crop/list.h"
 
 namespace arm_compute
 {
diff --git a/src/cpu/kernels/CpuActivationKernel.cpp b/src/cpu/kernels/CpuActivationKernel.cpp
index cfe732ab5f..d5112b4ba9 100644
--- a/src/cpu/kernels/CpuActivationKernel.cpp
+++ b/src/cpu/kernels/CpuActivationKernel.cpp
@@ -46,8 +46,23 @@ namespace
 static const std::vector<CpuActivationKernel::ActivationKernel> available_kernels =
 {
     {
+        "sve2_qu8_activation",
+        [](const DataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.isa.sve2; },
+        REGISTER_QASYMM8_SVE2(arm_compute::cpu::sve2_qasymm8_activation)
+    },
+    {
+        "sve2_qs8_activation",
+        [](const DataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED && data.isa.sve2; },
+        REGISTER_QASYMM8_SIGNED_SVE2(arm_compute::cpu::sve2_qasymm8_signed_activation)
+    },
+    {
+        "sve2_qs16_activation",
+        [](const DataTypeISASelectorData & data) { return data.dt == DataType::QSYMM16 && data.isa.sve2; },
+        REGISTER_QSYMM16_SVE2(arm_compute::cpu::sve2_qsymm16_activation)
+    },
+    {
         "sve_fp16_activation",
-        [](const DataTypeISASelectorData & data) { return data.dt == DataType::F16 && data.isa.sve; },
+        [](const DataTypeISASelectorData & data) { return data.dt == DataType::F16 && data.isa.sve && data.isa.fp16; },
         REGISTER_FP16_SVE(arm_compute::cpu::sve_fp16_activation)
     },
     {
@@ -66,21 +81,6 @@ static const std::vector<CpuActivationKernel::ActivationKernel> available_kernel
         REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_activation)
     },
     {
-        "sve2_qu8_activation",
-        [](const DataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.isa.sve2; },
-        REGISTER_QASYMM8_SVE2(arm_compute::cpu::sve2_qasymm8_activation)
-    },
-    {
-        "sve2_qs8_activation",
-        [](const DataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED && data.isa.sve2; },
-        REGISTER_QASYMM8_SIGNED_SVE2(arm_compute::cpu::sve2_qasymm8_signed_activation)
-    },
-    {
-        "sve2_qs16_activation",
-        [](const DataTypeISASelectorData & data) { return data.dt == DataType::QSYMM16 && data.isa.sve2; },
-        REGISTER_QSYMM16_SVE2(arm_compute::cpu::sve2_qsymm16_activation)
-    },
-    {
         "neon_qu8_activation",
         [](const DataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8; },
         REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_activation)
diff --git a/src/cpu/kernels/CpuAddKernel.cpp b/src/cpu/kernels/CpuAddKernel.cpp
index d06621fae0..e756effea9 100644
--- a/src/cpu/kernels/CpuAddKernel.cpp
+++ b/src/cpu/kernels/CpuAddKernel.cpp
@@ -79,7 +79,7 @@ static const std::vector<CpuAddKernel::AddKernel> available_kernels =
         "sve_fp16_add",
         [](const DataTypeISASelectorData & data)
         {
-            return (data.dt == DataType::F16) && data.isa.sve;
+            return (data.dt == DataType::F16) && data.isa.sve && data.isa.fp16;
         },
         REGISTER_FP16_SVE(arm_compute::cpu::add_fp16_sve)
     },
diff --git a/src/cpu/kernels/CpuPool2dKernel.cpp b/src/cpu/kernels/CpuPool2dKernel.cpp
index d0ca2d285d..8f04812b0c 100644
--- a/src/cpu/kernels/CpuPool2dKernel.cpp
+++ b/src/cpu/kernels/CpuPool2dKernel.cpp
@@ -64,13 +64,11 @@ static const std::vector<CpuPool2dKernel::PoolingKernel> available_kernels =
         [](const PoolDataTypeISASelectorData & data) { return ((data.dl == DataLayout::NHWC) && (data.dt == DataType::QASYMM8_SIGNED)); },
         REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::poolingMxN_qasymm8_signed_neon_nhwc)
     },
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
     {
         "neon_f16_nhwc_poolMxN",
-        [](const PoolDataTypeISASelectorData & data) { return ((data.dl == DataLayout::NHWC) && (data.dt == DataType::F16)); },
+        [](const PoolDataTypeISASelectorData & data) { return ((data.dl == DataLayout::NHWC) && (data.dt == DataType::F16)) && data.isa.fp16; },
         REGISTER_FP16_NEON(arm_compute::cpu::poolingMxN_fp16_neon_nhwc)
     },
-#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
     {
         "neon_fp32_nhwc_poolMxN",
         [](const PoolDataTypeISASelectorData & data) { return ((data.dl == DataLayout::NHWC) && (data.dt == DataType::F32)); },
@@ -107,7 +105,6 @@ static const std::vector<CpuPool2dKernel::PoolingKernel> available_kernels =
         [](const PoolDataTypeISASelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::QASYMM8_SIGNED)); },
         REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::poolingMxN_quantized_neon_nchw<int8_t>)
     },
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
     {
         "neon_fp16_nchw_pool2",
         [](const PoolDataTypeISASelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F16 && data.isa.fp16) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 2)); },
@@ -123,7 +120,6 @@ static const std::vector<CpuPool2dKernel::PoolingKernel> available_kernels =
         [](const PoolDataTypeISASelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F16 && data.isa.fp16)); },
         REGISTER_FP16_NEON(arm_compute::cpu::poolingMxN_fp16_neon_nchw)
     },
-#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
     {
         "neon_fp32_nchw_pool2",
         [](const PoolDataTypeISASelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F32) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 2)); },
diff --git a/src/cpu/kernels/CpuScaleKernel.cpp b/src/cpu/kernels/CpuScaleKernel.cpp
index 60564a97dd..e230dfa938 100644
--- a/src/cpu/kernels/CpuScaleKernel.cpp
+++ b/src/cpu/kernels/CpuScaleKernel.cpp
@@ -50,10 +50,9 @@ namespace
 {
 static const std::vector<CpuScaleKernel::ScaleKernel> available_kernels =
 {
-#if defined(ARM_COMPUTE_ENABLE_SVE)
     {
         "sve_fp16_scale",
-        [](const DataTypeISASelectorData & data) { return data.dt == DataType::F16 && data.isa.sve; },
+        [](const DataTypeISASelectorData & data) { return data.dt == DataType::F16 && data.isa.sve && data.isa.fp16; },
         REGISTER_FP16_SVE(arm_compute::cpu::fp16_sve_scale)
     },
     {
@@ -81,15 +80,11 @@ static const std::vector<CpuScaleKernel::ScaleKernel> available_kernels =
         [](const DataTypeISASelectorData & data) { return data.dt == DataType::S16 && data.isa.sve; },
         REGISTER_INTEGER_SVE(arm_compute::cpu::s16_sve_scale)
     },
-#endif /* defined(ARM_COMPUTE_ENABLE_SVE) */
-#if defined(ARM_COMPUTE_ENABLE_NEON)
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
     {
         "neon_fp16_scale",
         [](const DataTypeISASelectorData & data) { return data.dt == DataType::F16 && data.isa.fp16; },
         REGISTER_FP16_NEON(arm_compute::cpu::common_neon_scale<float16_t>)
     },
-#endif /* !defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
     {
         "neon_fp32_scale",
         [](const DataTypeISASelectorData & data) { return data.dt == DataType::F32; },
@@ -115,7 +110,6 @@ static const std::vector<CpuScaleKernel::ScaleKernel> available_kernels =
         [](const DataTypeISASelectorData & data) { return data.dt == DataType::S16; },
         REGISTER_INTEGER_NEON(arm_compute::cpu::s16_neon_scale)
     },
-#endif /* defined(ARM_COMPUTE_ENABLE_NEON) */
 };
 
 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy,
diff --git a/src/cpu/kernels/CpuSubKernel.cpp b/src/cpu/kernels/CpuSubKernel.cpp
index c12feb4331..c55d11e899 100644
--- a/src/cpu/kernels/CpuSubKernel.cpp
+++ b/src/cpu/kernels/CpuSubKernel.cpp
@@ -46,13 +46,11 @@ static const std::vector<CpuSubKernel::SubKernel> available_kernels =
         [](const DataTypeISASelectorData & data) { return (data.dt == DataType::F32); },
         REGISTER_FP32_NEON(arm_compute::cpu::sub_same_neon<float>)
     },
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
     {
         "neon_fp16_sub",
         [](const DataTypeISASelectorData & data) { return (data.dt == DataType::F16) && data.isa.fp16; },
         REGISTER_FP16_NEON(arm_compute::cpu::sub_same_neon<float16_t>)
     },
-#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
     {
         "neon_u8_sub",
         [](const DataTypeISASelectorData & data) { return (data.dt == DataType::U8); },
diff --git a/src/cpu/kernels/crop/generic/neon/integer.cpp b/src/cpu/kernels/crop/generic/neon/integer.cpp
index 7dbf0a7f5e..ebf2c1fbd3 100644
--- a/src/cpu/kernels/crop/generic/neon/integer.cpp
+++ b/src/cpu/kernels/crop/generic/neon/integer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,7 +23,7 @@
  */
 
 #include "src/cpu/kernels/crop/generic/neon/impl.h"
-#include "src/cpu/kernels/crop/generic/neon/list.h"
+#include "src/cpu/kernels/crop/list.h"
 
 namespace arm_compute
 {
diff --git a/src/cpu/kernels/crop/generic/neon/list.h b/src/cpu/kernels/crop/list.h
index f33049304d..a6b83215ae 100644
--- a/src/cpu/kernels/crop/generic/neon/list.h
+++ b/src/cpu/kernels/crop/list.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/src/cpu/kernels/elementwise_binary/generic/neon/impl.h b/src/cpu/kernels/elementwise_binary/generic/neon/impl.h
index ead54ab14e..98b154e8fd 100644
--- a/src/cpu/kernels/elementwise_binary/generic/neon/impl.h
+++ b/src/cpu/kernels/elementwise_binary/generic/neon/impl.h
@@ -69,6 +69,7 @@ typename VectorType::type elementwise_arithm_op(const typename VectorType::type
 
     return res;
 }
+
 template <ArithmeticOperation op, typename ScalarType, typename VectorType>
 typename VectorType::type elementwise_arithm_op_broadcast(const typename VectorType::type &a, const ScalarType &broadcast_value, const bool reorder)
 {
diff --git a/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp b/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp
new file mode 100644
index 0000000000..30caa4ebeb
--- /dev/null
+++ b/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp
@@ -0,0 +1,115 @@
+/*
+ * Copyright (c) 2018-2022 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 "src/cpu/kernels/elementwise_unary/generic/neon/impl.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <typename ScalarType>
+inline ScalarType elementwise_op_scalar_imp(ElementWiseUnary op, const ScalarType &a)
+{
+    switch(op)
+    {
+        case ElementWiseUnary::RSQRT:
+            return 1 / sqrt(a);
+        case ElementWiseUnary::EXP:
+            return std::exp(a);
+        case ElementWiseUnary::NEG:
+            return -a;
+        case ElementWiseUnary::LOG:
+            return std::log(a);
+        case ElementWiseUnary::ABS:
+            return std::abs(a);
+        case ElementWiseUnary::ROUND:
+            return support::cpp11::nearbyint(a);
+        case ElementWiseUnary::SIN:
+            return std::sin(a);
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+}
+
+template <typename ScalarType, typename VectorType>
+inline VectorType elementwise_op_imp(ElementWiseUnary op, const VectorType &a)
+{
+    switch(op)
+    {
+        case ElementWiseUnary::RSQRT:
+            return wrapper::vinvsqrt(a);
+        case ElementWiseUnary::EXP:
+            return wrapper::vexpq(a);
+        case ElementWiseUnary::NEG:
+            return wrapper::vneg(a);
+        case ElementWiseUnary::LOG:
+            return wrapper::vlog(a);
+        case ElementWiseUnary::ABS:
+            return wrapper::vabs(a);
+        case ElementWiseUnary::ROUND:
+            return wrapper::vround(a);
+        case ElementWiseUnary::SIN:
+            return wrapper::vsin(a);
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+}
+
+template <typename ScalarType>
+void elementwise_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
+{
+    const int  window_step_x  = 16 / sizeof(ScalarType);
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    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;
+        for(; x <= window_end_x - window_step_x; x += window_step_x)
+        {
+            wrapper::vstore(output_ptr + x, elementwise_op_imp<ScalarType>(op, wrapper::vloadq(input_ptr + x)));
+        }
+        for(; x < window_end_x; ++x)
+        {
+            *(output_ptr + x) = elementwise_op_scalar_imp(op, *(input_ptr + x));
+        }
+    },
+    input, output);
+}
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+template void elementwise_op<__fp16>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
+#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+template void elementwise_op<float>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
+template void elementwise_op<int32_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
+
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/cpu/kernels/elementwise_unary/generic/neon/impl.h b/src/cpu/kernels/elementwise_unary/generic/neon/impl.h
index fd930ae7cf..66b8b5fe45 100644
--- a/src/cpu/kernels/elementwise_unary/generic/neon/impl.h
+++ b/src/cpu/kernels/elementwise_unary/generic/neon/impl.h
@@ -24,6 +24,7 @@
 #ifndef SRC_CORE_NEON_KERNELS_ELEMENTWISE_UNARY_LIST_H
 #define SRC_CORE_NEON_KERNELS_ELEMENTWISE_UNARY_LIST_H
 
+#include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Types.h"
 #include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
 
@@ -32,83 +33,7 @@ namespace arm_compute
 namespace cpu
 {
 template <typename ScalarType>
-inline ScalarType elementwise_op_scalar_imp(ElementWiseUnary op, const ScalarType &a)
-{
-    switch(op)
-    {
-        case ElementWiseUnary::RSQRT:
-            return 1 / sqrt(a);
-        case ElementWiseUnary::EXP:
-            return std::exp(a);
-        case ElementWiseUnary::NEG:
-            return -a;
-        case ElementWiseUnary::LOG:
-            return std::log(a);
-        case ElementWiseUnary::ABS:
-            return std::abs(a);
-        case ElementWiseUnary::ROUND:
-            return support::cpp11::nearbyint(a);
-        case ElementWiseUnary::SIN:
-            return std::sin(a);
-        default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
-    }
-}
-
-template <typename ScalarType, typename VectorType>
-inline VectorType elementwise_op_imp(ElementWiseUnary op, const VectorType &a)
-{
-    switch(op)
-    {
-        case ElementWiseUnary::RSQRT:
-            return wrapper::vinvsqrt(a);
-        case ElementWiseUnary::EXP:
-            return wrapper::vexpq(a);
-        case ElementWiseUnary::NEG:
-            return wrapper::vneg(a);
-        case ElementWiseUnary::LOG:
-            return wrapper::vlog(a);
-        case ElementWiseUnary::ABS:
-            return wrapper::vabs(a);
-        case ElementWiseUnary::ROUND:
-            return wrapper::vround(a);
-        case ElementWiseUnary::SIN:
-            return wrapper::vsin(a);
-        default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
-    }
-}
-
-template <typename ScalarType>
-void elementwise_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
-{
-    const int  window_step_x  = 16 / sizeof(ScalarType);
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
-
-    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;
-        for(; x <= window_end_x - window_step_x; x += window_step_x)
-        {
-            wrapper::vstore(output_ptr + x, elementwise_op_imp<ScalarType>(op, wrapper::vloadq(input_ptr + x)));
-        }
-        for(; x < window_end_x; ++x)
-        {
-            *(output_ptr + x) = elementwise_op_scalar_imp(op, *(input_ptr + x));
-        }
-    },
-    input, output);
-}
+void elementwise_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
 
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/cpu/kernels/softmax/generic/neon/impl.cpp b/src/cpu/kernels/softmax/generic/neon/impl.cpp
new file mode 100644
index 0000000000..5654bb52ca
--- /dev/null
+++ b/src/cpu/kernels/softmax/generic/neon/impl.cpp
@@ -0,0 +1,399 @@
+/*
+ * Copyright (c) 2021-2022 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 "src/cpu/kernels/softmax/generic/neon/impl.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "support/SaturateCast.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <typename T>
+void neon_logits_1d_max(const ITensor *in, ITensor *out, const Window &window)
+{
+    /** SIMD vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
+
+    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());
+
+    Window win{ window };
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator input(in, win);
+    Iterator output(out, win);
+
+    const int sum_stages = log2(window_step_x / 2);
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        // Get pointers
+        const auto in_ptr  = reinterpret_cast<const T *>(input.ptr());
+        const auto out_ptr = reinterpret_cast<T *>(output.ptr());
+
+        // Init max value
+        auto vec_max = wrapper::vdup_n(support::cpp11::lowest<T>(), ExactTagType{});
+        int  x       = window_start_x;
+
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            const auto current_value = wrapper::vloadq(in_ptr + x);
+            vec_max                  = wrapper::vmax(vec_max, current_value);
+        }
+        auto carry_max = wrapper::vpmax(wrapper::vgethigh(vec_max), wrapper::vgetlow(vec_max));
+
+        for(int i = 0; i < sum_stages; ++i)
+        {
+            carry_max = wrapper::vpmax(carry_max, carry_max);
+        }
+        T max_val = wrapper::vgetlane(carry_max, 0);
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            max_val = *(in_ptr + x) > max_val ? *(in_ptr + x) : max_val;
+        }
+
+        *out_ptr = max_val;
+    },
+    input, output);
+}
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+template void neon_logits_1d_max<float16_t>(const ITensor *in, ITensor *out, const Window &window);
+#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+template void neon_logits_1d_max<float>(const ITensor *in, ITensor *out, const Window &window);
+template void neon_logits_1d_max<qasymm8_signed_t>(const ITensor *in, ITensor *out, const Window &window);
+template void neon_logits_1d_max<qasymm8_t>(const ITensor *in, ITensor *out, const Window &window);
+
+template <typename T>
+void neon_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp,
+                                      ITensor *out, float beta, bool is_log, const Window &window)
+{
+    static_assert(std::is_same<T, qasymm8_t>::value
+                  || std::is_same<T, qasymm8_signed_t>::value,
+                  "quantized type should be either qasymm8_t or qasymm8_signed_t.");
+
+    const int start_x     = in->info()->valid_region().anchor.x();
+    const int input_width = in->info()->valid_region().shape.x();
+
+    const float scale_beta     = -beta * in->info()->quantization_info().uniform().scale;
+    const auto  scale_beta_vec = vdupq_n_f32(scale_beta);
+
+    Iterator      in_it(in, window);
+    Iterator      max_it(max, window);
+    Iterator      out_it(out, window);
+    constexpr int vec_size = 16;
+
+    execute_window_loop(window, [&](const Coordinates &)
+    {
+        /* Get pointers */
+        const auto in_ptr  = reinterpret_cast<const T *>(in_it.ptr()) + start_x;
+        const auto out_ptr = reinterpret_cast<T *>(out_it.ptr()) + start_x;
+        const auto tmp_ptr = reinterpret_cast<float *>(tmp);
+
+        float sum{};
+        float sum_inversed{};
+
+        /* Compute exponentials and sum */
+        {
+            /* Get max value */
+            const auto max_val = *reinterpret_cast<const T *>(max_it.ptr());
+            const auto vec_max = wrapper::vdup_n(max_val, wrapper::traits::vector_128_tag{});
+
+            /* Init sum to zero */
+            float32x4x4_t vec_sum =
+            {
+                vdupq_n_f32(0.f),
+                vdupq_n_f32(0.f),
+                vdupq_n_f32(0.f),
+                vdupq_n_f32(0.f),
+            };
+
+            /* Loop over row and compute exponentials and sum */
+            int x = 0;
+            for(; x <= (input_width - vec_size); x += vec_size)
+            {
+                auto vec_elements     = wrapper::vloadq(in_ptr + x);
+                vec_elements          = wrapper::vqsub(vec_max, vec_elements);
+                auto vec_elements_flt = convert_int_to_float<float32x4x4_t>(vec_elements);
+
+                if(is_log)
+                {
+                    vec_elements_flt.val[0] = vmulq_f32(vec_elements_flt.val[0], scale_beta_vec);
+                    vec_elements_flt.val[1] = vmulq_f32(vec_elements_flt.val[1], scale_beta_vec);
+                    vec_elements_flt.val[2] = vmulq_f32(vec_elements_flt.val[2], scale_beta_vec);
+                    vec_elements_flt.val[3] = vmulq_f32(vec_elements_flt.val[3], scale_beta_vec);
+                    vec_sum.val[0]          = vaddq_f32(vec_sum.val[0], vexpq_f32(vec_elements_flt.val[0]));
+                    vec_sum.val[1]          = vaddq_f32(vec_sum.val[1], vexpq_f32(vec_elements_flt.val[1]));
+                    vec_sum.val[2]          = vaddq_f32(vec_sum.val[2], vexpq_f32(vec_elements_flt.val[2]));
+                    vec_sum.val[3]          = vaddq_f32(vec_sum.val[3], vexpq_f32(vec_elements_flt.val[3]));
+                }
+                else
+                {
+                    vec_elements_flt.val[0] = vexpq_f32(vmulq_f32(vec_elements_flt.val[0], scale_beta_vec));
+                    vec_elements_flt.val[1] = vexpq_f32(vmulq_f32(vec_elements_flt.val[1], scale_beta_vec));
+                    vec_elements_flt.val[2] = vexpq_f32(vmulq_f32(vec_elements_flt.val[2], scale_beta_vec));
+                    vec_elements_flt.val[3] = vexpq_f32(vmulq_f32(vec_elements_flt.val[3], scale_beta_vec));
+                    vec_sum.val[0]          = vaddq_f32(vec_sum.val[0], vec_elements_flt.val[0]);
+                    vec_sum.val[1]          = vaddq_f32(vec_sum.val[1], vec_elements_flt.val[1]);
+                    vec_sum.val[2]          = vaddq_f32(vec_sum.val[2], vec_elements_flt.val[2]);
+                    vec_sum.val[3]          = vaddq_f32(vec_sum.val[3], vec_elements_flt.val[3]);
+                }
+
+                vst4q_f32(tmp_ptr + x, vec_elements_flt);
+            }
+
+            /* Reduce sum */
+            const auto sum_16_byte = vaddq_f32(vaddq_f32(vec_sum.val[0], vec_sum.val[1]), vaddq_f32(vec_sum.val[2], vec_sum.val[3]));
+            auto       sum_res     = vpadd_f32(vget_high_f32(sum_16_byte), vget_low_f32(sum_16_byte));
+            sum_res                = vpadd_f32(sum_res, sum_res);
+            sum                    = wrapper::vgetlane(sum_res, 0);
+
+            /* Run remaining elements */
+            for(; x < input_width; ++x)
+            {
+                float element{};
+                if(is_log)
+                {
+                    element = (max_val - in_ptr[x]) * scale_beta;
+                    sum += std::exp(element);
+                }
+                else
+                {
+                    element = std::exp((max_val - in_ptr[x]) * scale_beta);
+                    sum += element;
+                }
+
+                tmp_ptr[x] = element;
+            }
+
+            if(!is_log)
+            {
+                sum_inversed = 256.f / sum;
+            }
+            else
+            {
+                sum = std::log(sum);
+            }
+        }
+
+        /* Normalize exponentials */
+        {
+            constexpr bool is_qasymm8_signed = std::is_same<T, qasymm8_signed_t>::value;
+            /* Loop over row and compute softmax */
+            int x = 0;
+            for(; x <= (input_width - vec_size); x += vec_size)
+            {
+                using int_vec_type   = wrapper::traits::neon_vector_t<T, 16>;
+                float32x4x4_t vec_in = vld4q_f32(tmp_ptr + x);
+                int_vec_type  normalized_value{};
+                if(is_log)
+                {
+                    const float32x4x4_t sub =
+                    {
+                        vsubq_f32(vec_in.val[0], vdupq_n_f32(sum)),
+                        vsubq_f32(vec_in.val[1], vdupq_n_f32(sum)),
+                        vsubq_f32(vec_in.val[2], vdupq_n_f32(sum)),
+                        vsubq_f32(vec_in.val[3], vdupq_n_f32(sum)),
+                    };
+                    normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(sub);
+                }
+                else
+                {
+                    float32x4x4_t mul =
+                    {
+                        vmulq_f32(vec_in.val[0], vdupq_n_f32(sum_inversed)),
+                        vmulq_f32(vec_in.val[1], vdupq_n_f32(sum_inversed)),
+                        vmulq_f32(vec_in.val[2], vdupq_n_f32(sum_inversed)),
+                        vmulq_f32(vec_in.val[3], vdupq_n_f32(sum_inversed)),
+                    };
+
+                    if(is_qasymm8_signed)
+                    {
+                        const auto offset_vec = wrapper::vdup_n(128.f, wrapper::traits::vector_128_tag{});
+                        mul.val[0]            = wrapper::vsub(mul.val[0], offset_vec);
+                        mul.val[1]            = wrapper::vsub(mul.val[1], offset_vec);
+                        mul.val[2]            = wrapper::vsub(mul.val[2], offset_vec);
+                        mul.val[3]            = wrapper::vsub(mul.val[3], offset_vec);
+                    }
+
+                    normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(mul);
+                }
+                wrapper::vstore(out_ptr + x, normalized_value);
+            }
+            /* Run remaining elements */
+            for(; x < input_width; ++x)
+            {
+                if(is_log)
+                {
+                    out_ptr[x] = utils::cast::saturate_cast<T>(tmp_ptr[x] - sum);
+                }
+                else
+                {
+                    out_ptr[x] = utils::cast::saturate_cast<T>((tmp_ptr[x] * sum_inversed) - (is_qasymm8_signed ? 128.f : 0));
+                }
+            }
+        }
+    },
+    in_it, max_it, out_it);
+}
+
+template void neon_softmax_logits_1d_quantized<qasymm8_signed_t>(const ITensor *in, const ITensor *max, void *const tmp,
+                                                                 ITensor *out, float beta, bool is_log, const Window &window);
+template void neon_softmax_logits_1d_quantized<qasymm8_t>(const ITensor *in, const ITensor *max, void *const tmp,
+                                                          ITensor *out, float beta, bool is_log, const Window &window);
+template <typename T>
+void neon_softmax_logits_1d_float(const ITensor *in, const ITensor *max, void *const tmp,
+                                  ITensor *out, const float beta, bool is_log, const Window &window)
+{
+    const int start_x     = in->info()->valid_region().anchor.x();
+    const int input_width = in->info()->valid_region().shape.x();
+
+    Iterator in_it(in, window);
+    Iterator max_it(max, window);
+    Iterator out_it(out, window);
+
+    /** SIMD vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
+
+    constexpr int vec_size   = 16 / sizeof(T);
+    const int     sum_stages = log2(vec_size / 2);
+
+    execute_window_loop(window, [&](const Coordinates &)
+    {
+        /* Get pointers */
+        const auto in_ptr  = reinterpret_cast<const T *>(in_it.ptr()) + start_x;
+        const auto out_ptr = reinterpret_cast<T *>(out_it.ptr()) + start_x;
+        const auto tmp_ptr = reinterpret_cast<T *>(tmp);
+
+        T sum{};
+        T sum_inversed{};
+
+        /* Compute exponentials and sum */
+        {
+            /* Get max value */
+            const auto max_val = *reinterpret_cast<const T *>(max_it.ptr());
+            const auto vec_max = wrapper::vdup_n(max_val, ExactTagType{});
+
+            /* Init sum to zero */
+            auto vec_sum = wrapper::vdup_n(static_cast<T>(0), ExactTagType{});
+
+            /* Loop over row and compute exponentials and sum */
+            int x = 0;
+            for(; x <= (input_width - vec_size); x += vec_size)
+            {
+                auto vec_elements = wrapper::vloadq(in_ptr + x);
+                vec_elements      = wrapper::vsub(vec_elements, vec_max);
+                if(is_log)
+                {
+                    vec_elements = wrapper::vmul(vec_elements, wrapper::vdup_n(static_cast<T>(beta), ExactTagType{}));
+                    vec_sum      = wrapper::vadd(vec_sum, wrapper::vexpq(vec_elements));
+                }
+                else
+                {
+                    vec_elements = wrapper::vexpq(wrapper::vmul(vec_elements, wrapper::vdup_n(static_cast<T>(beta), ExactTagType{})));
+                    vec_sum      = wrapper::vadd(vec_sum, vec_elements);
+                }
+                wrapper::vstore(tmp_ptr + x, vec_elements);
+            }
+
+            /* Reduce sum */
+            auto sum_res = wrapper::vpadd(wrapper::vgethigh(vec_sum), wrapper::vgetlow(vec_sum));
+            for(int i = 0; i < sum_stages; ++i)
+            {
+                sum_res = wrapper::vpadd(sum_res, sum_res);
+            }
+            sum = wrapper::vgetlane(sum_res, 0);
+
+            /* Run remaining elements */
+            for(; x < input_width; ++x)
+            {
+                T element{};
+
+                if(is_log)
+                {
+                    element = (in_ptr[x] - max_val) * beta;
+                    sum += std::exp(element);
+                }
+                else
+                {
+                    element = std::exp((in_ptr[x] - max_val) * beta);
+                    sum += element;
+                }
+                tmp_ptr[x] = element;
+            }
+
+            if(!is_log)
+            {
+                sum_inversed = T(1) / sum;
+            }
+            else
+            {
+                sum = static_cast<T>(std::log(sum));
+            }
+        }
+
+        /* Normalize exponentials */
+        {
+            /* Loop over row and compute softmax */
+            int x = 0;
+            for(; x <= (input_width - vec_size); x += vec_size)
+            {
+                auto vec_in           = wrapper::vloadq(tmp_ptr + x);
+                auto normalized_value = wrapper::vdup_n(static_cast<T>(0), ExactTagType{});
+                if(is_log)
+                {
+                    normalized_value = wrapper::vsub(vec_in, wrapper::vdup_n(static_cast<T>(sum), ExactTagType{}));
+                }
+                else
+                {
+                    normalized_value = wrapper::vmul(vec_in, wrapper::vdup_n(static_cast<T>(sum_inversed), ExactTagType{}));
+                }
+                wrapper::vstore(out_ptr + x, normalized_value);
+            }
+            /* Run remaining elements */
+            for(; x < input_width; ++x)
+            {
+                if(is_log)
+                {
+                    out_ptr[x] = tmp_ptr[x] - sum;
+                }
+                else
+                {
+                    out_ptr[x] = tmp_ptr[x] * sum_inversed;
+                }
+            }
+        }
+    },
+    in_it, max_it, out_it);
+}
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+template void neon_softmax_logits_1d_float<float16_t>(const ITensor *in, const ITensor *max, void *const tmp,
+                                                      ITensor *out, const float beta, bool is_log, const Window &window);
+#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+template void neon_softmax_logits_1d_float<float>(const ITensor *in, const ITensor *max, void *const tmp,
+                                                  ITensor *out, const float beta, bool is_log, const Window &window);
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/cpu/kernels/softmax/generic/neon/impl.h b/src/cpu/kernels/softmax/generic/neon/impl.h
index 325e127f3b..6ca659919a 100644
--- a/src/cpu/kernels/softmax/generic/neon/impl.h
+++ b/src/cpu/kernels/softmax/generic/neon/impl.h
@@ -24,363 +24,22 @@
 #ifndef SRC_CORE_NEON_KERNELS_SOFTMAX_IMPL_H
 #define SRC_CORE_NEON_KERNELS_SOFTMAX_IMPL_H
 
-#include "src/core/NEON/NEMath.h"
-#include "src/core/NEON/wrapper/wrapper.h"
-#include "support/SaturateCast.h"
+#include "arm_compute/core/Helpers.h"
 
 namespace arm_compute
 {
 namespace cpu
 {
 template <typename T>
-void neon_logits_1d_max(const ITensor *in, ITensor *out, const Window &window)
-{
-    /** SIMD vector tag type. */
-    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
-
-    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());
-
-    Window win{ window };
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    Iterator input(in, win);
-    Iterator output(out, win);
-
-    const int sum_stages = log2(window_step_x / 2);
-    execute_window_loop(win, [&](const Coordinates &)
-    {
-        // Get pointers
-        const auto in_ptr  = reinterpret_cast<const T *>(input.ptr());
-        const auto out_ptr = reinterpret_cast<T *>(output.ptr());
-
-        // Init max value
-        auto vec_max = wrapper::vdup_n(support::cpp11::lowest<T>(), ExactTagType{});
-        int  x       = window_start_x;
-
-        for(; x <= (window_end_x - window_step_x); x += window_step_x)
-        {
-            const auto current_value = wrapper::vloadq(in_ptr + x);
-            vec_max                  = wrapper::vmax(vec_max, current_value);
-        }
-        auto carry_max = wrapper::vpmax(wrapper::vgethigh(vec_max), wrapper::vgetlow(vec_max));
-
-        for(int i = 0; i < sum_stages; ++i)
-        {
-            carry_max = wrapper::vpmax(carry_max, carry_max);
-        }
-        T max_val = wrapper::vgetlane(carry_max, 0);
-
-        // Compute left-over elements
-        for(; x < window_end_x; ++x)
-        {
-            max_val = *(in_ptr + x) > max_val ? *(in_ptr + x) : max_val;
-        }
-
-        *out_ptr = max_val;
-    },
-    input, output);
-}
+void neon_logits_1d_max(const ITensor *in, ITensor *out, const Window &window);
 
 template <typename T>
 void neon_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp,
-                                      ITensor *out, float beta, bool is_log, const Window &window)
-{
-    static_assert(std::is_same<T, qasymm8_t>::value
-                  || std::is_same<T, qasymm8_signed_t>::value,
-                  "quantized type should be either qasymm8_t or qasymm8_signed_t.");
-
-    const int start_x     = in->info()->valid_region().anchor.x();
-    const int input_width = in->info()->valid_region().shape.x();
-
-    const float scale_beta     = -beta * in->info()->quantization_info().uniform().scale;
-    const auto  scale_beta_vec = vdupq_n_f32(scale_beta);
-
-    Iterator      in_it(in, window);
-    Iterator      max_it(max, window);
-    Iterator      out_it(out, window);
-    constexpr int vec_size = 16;
-
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        /* Get pointers */
-        const auto in_ptr  = reinterpret_cast<const T *>(in_it.ptr()) + start_x;
-        const auto out_ptr = reinterpret_cast<T *>(out_it.ptr()) + start_x;
-        const auto tmp_ptr = reinterpret_cast<float *>(tmp);
-
-        float sum{};
-        float sum_inversed{};
-
-        /* Compute exponentials and sum */
-        {
-            /* Get max value */
-            const auto max_val = *reinterpret_cast<const T *>(max_it.ptr());
-            const auto vec_max = wrapper::vdup_n(max_val, wrapper::traits::vector_128_tag{});
-
-            /* Init sum to zero */
-            float32x4x4_t vec_sum =
-            {
-                vdupq_n_f32(0.f),
-                vdupq_n_f32(0.f),
-                vdupq_n_f32(0.f),
-                vdupq_n_f32(0.f),
-            };
-
-            /* Loop over row and compute exponentials and sum */
-            int x = 0;
-            for(; x <= (input_width - vec_size); x += vec_size)
-            {
-                auto vec_elements     = wrapper::vloadq(in_ptr + x);
-                vec_elements          = wrapper::vqsub(vec_max, vec_elements);
-                auto vec_elements_flt = convert_int_to_float<float32x4x4_t>(vec_elements);
-
-                if(is_log)
-                {
-                    vec_elements_flt.val[0] = vmulq_f32(vec_elements_flt.val[0], scale_beta_vec);
-                    vec_elements_flt.val[1] = vmulq_f32(vec_elements_flt.val[1], scale_beta_vec);
-                    vec_elements_flt.val[2] = vmulq_f32(vec_elements_flt.val[2], scale_beta_vec);
-                    vec_elements_flt.val[3] = vmulq_f32(vec_elements_flt.val[3], scale_beta_vec);
-                    vec_sum.val[0]          = vaddq_f32(vec_sum.val[0], vexpq_f32(vec_elements_flt.val[0]));
-                    vec_sum.val[1]          = vaddq_f32(vec_sum.val[1], vexpq_f32(vec_elements_flt.val[1]));
-                    vec_sum.val[2]          = vaddq_f32(vec_sum.val[2], vexpq_f32(vec_elements_flt.val[2]));
-                    vec_sum.val[3]          = vaddq_f32(vec_sum.val[3], vexpq_f32(vec_elements_flt.val[3]));
-                }
-                else
-                {
-                    vec_elements_flt.val[0] = vexpq_f32(vmulq_f32(vec_elements_flt.val[0], scale_beta_vec));
-                    vec_elements_flt.val[1] = vexpq_f32(vmulq_f32(vec_elements_flt.val[1], scale_beta_vec));
-                    vec_elements_flt.val[2] = vexpq_f32(vmulq_f32(vec_elements_flt.val[2], scale_beta_vec));
-                    vec_elements_flt.val[3] = vexpq_f32(vmulq_f32(vec_elements_flt.val[3], scale_beta_vec));
-                    vec_sum.val[0]          = vaddq_f32(vec_sum.val[0], vec_elements_flt.val[0]);
-                    vec_sum.val[1]          = vaddq_f32(vec_sum.val[1], vec_elements_flt.val[1]);
-                    vec_sum.val[2]          = vaddq_f32(vec_sum.val[2], vec_elements_flt.val[2]);
-                    vec_sum.val[3]          = vaddq_f32(vec_sum.val[3], vec_elements_flt.val[3]);
-                }
-
-                vst4q_f32(tmp_ptr + x, vec_elements_flt);
-            }
-
-            /* Reduce sum */
-            const auto sum_16_byte = vaddq_f32(vaddq_f32(vec_sum.val[0], vec_sum.val[1]), vaddq_f32(vec_sum.val[2], vec_sum.val[3]));
-            auto       sum_res     = vpadd_f32(vget_high_f32(sum_16_byte), vget_low_f32(sum_16_byte));
-            sum_res                = vpadd_f32(sum_res, sum_res);
-            sum                    = wrapper::vgetlane(sum_res, 0);
-
-            /* Run remaining elements */
-            for(; x < input_width; ++x)
-            {
-                float element{};
-                if(is_log)
-                {
-                    element = (max_val - in_ptr[x]) * scale_beta;
-                    sum += std::exp(element);
-                }
-                else
-                {
-                    element = std::exp((max_val - in_ptr[x]) * scale_beta);
-                    sum += element;
-                }
-
-                tmp_ptr[x] = element;
-            }
-
-            if(!is_log)
-            {
-                sum_inversed = 256.f / sum;
-            }
-            else
-            {
-                sum = std::log(sum);
-            }
-        }
-
-        /* Normalize exponentials */
-        {
-            constexpr bool is_qasymm8_signed = std::is_same<T, qasymm8_signed_t>::value;
-            /* Loop over row and compute softmax */
-            int x = 0;
-            for(; x <= (input_width - vec_size); x += vec_size)
-            {
-                using int_vec_type   = wrapper::traits::neon_vector_t<T, 16>;
-                float32x4x4_t vec_in = vld4q_f32(tmp_ptr + x);
-                int_vec_type  normalized_value{};
-                if(is_log)
-                {
-                    const float32x4x4_t sub =
-                    {
-                        vsubq_f32(vec_in.val[0], vdupq_n_f32(sum)),
-                        vsubq_f32(vec_in.val[1], vdupq_n_f32(sum)),
-                        vsubq_f32(vec_in.val[2], vdupq_n_f32(sum)),
-                        vsubq_f32(vec_in.val[3], vdupq_n_f32(sum)),
-                    };
-                    normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(sub);
-                }
-                else
-                {
-                    float32x4x4_t mul =
-                    {
-                        vmulq_f32(vec_in.val[0], vdupq_n_f32(sum_inversed)),
-                        vmulq_f32(vec_in.val[1], vdupq_n_f32(sum_inversed)),
-                        vmulq_f32(vec_in.val[2], vdupq_n_f32(sum_inversed)),
-                        vmulq_f32(vec_in.val[3], vdupq_n_f32(sum_inversed)),
-                    };
-
-                    if(is_qasymm8_signed)
-                    {
-                        const auto offset_vec = wrapper::vdup_n(128.f, wrapper::traits::vector_128_tag{});
-                        mul.val[0]            = wrapper::vsub(mul.val[0], offset_vec);
-                        mul.val[1]            = wrapper::vsub(mul.val[1], offset_vec);
-                        mul.val[2]            = wrapper::vsub(mul.val[2], offset_vec);
-                        mul.val[3]            = wrapper::vsub(mul.val[3], offset_vec);
-                    }
-
-                    normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(mul);
-                }
-                wrapper::vstore(out_ptr + x, normalized_value);
-            }
-            /* Run remaining elements */
-            for(; x < input_width; ++x)
-            {
-                if(is_log)
-                {
-                    out_ptr[x] = utils::cast::saturate_cast<T>(tmp_ptr[x] - sum);
-                }
-                else
-                {
-                    out_ptr[x] = utils::cast::saturate_cast<T>((tmp_ptr[x] * sum_inversed) - (is_qasymm8_signed ? 128.f : 0));
-                }
-            }
-        }
-    },
-    in_it, max_it, out_it);
-}
+                                      ITensor *out, float beta, bool is_log, const Window &window);
 
 template <typename T>
 void neon_softmax_logits_1d_float(const ITensor *in, const ITensor *max, void *const tmp,
-                                  ITensor *out, const float beta, bool is_log, const Window &window)
-{
-    const int start_x     = in->info()->valid_region().anchor.x();
-    const int input_width = in->info()->valid_region().shape.x();
-
-    Iterator in_it(in, window);
-    Iterator max_it(max, window);
-    Iterator out_it(out, window);
-
-    /** SIMD vector tag type. */
-    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
-
-    constexpr int vec_size   = 16 / sizeof(T);
-    const int     sum_stages = log2(vec_size / 2);
-
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        /* Get pointers */
-        const auto in_ptr  = reinterpret_cast<const T *>(in_it.ptr()) + start_x;
-        const auto out_ptr = reinterpret_cast<T *>(out_it.ptr()) + start_x;
-        const auto tmp_ptr = reinterpret_cast<T *>(tmp);
-
-        T sum{};
-        T sum_inversed{};
-
-        /* Compute exponentials and sum */
-        {
-            /* Get max value */
-            const auto max_val = *reinterpret_cast<const T *>(max_it.ptr());
-            const auto vec_max = wrapper::vdup_n(max_val, ExactTagType{});
-
-            /* Init sum to zero */
-            auto vec_sum = wrapper::vdup_n(static_cast<T>(0), ExactTagType{});
-
-            /* Loop over row and compute exponentials and sum */
-            int x = 0;
-            for(; x <= (input_width - vec_size); x += vec_size)
-            {
-                auto vec_elements = wrapper::vloadq(in_ptr + x);
-                vec_elements      = wrapper::vsub(vec_elements, vec_max);
-                if(is_log)
-                {
-                    vec_elements = wrapper::vmul(vec_elements, wrapper::vdup_n(static_cast<T>(beta), ExactTagType{}));
-                    vec_sum      = wrapper::vadd(vec_sum, wrapper::vexpq(vec_elements));
-                }
-                else
-                {
-                    vec_elements = wrapper::vexpq(wrapper::vmul(vec_elements, wrapper::vdup_n(static_cast<T>(beta), ExactTagType{})));
-                    vec_sum      = wrapper::vadd(vec_sum, vec_elements);
-                }
-                wrapper::vstore(tmp_ptr + x, vec_elements);
-            }
-
-            /* Reduce sum */
-            auto sum_res = wrapper::vpadd(wrapper::vgethigh(vec_sum), wrapper::vgetlow(vec_sum));
-            for(int i = 0; i < sum_stages; ++i)
-            {
-                sum_res = wrapper::vpadd(sum_res, sum_res);
-            }
-            sum = wrapper::vgetlane(sum_res, 0);
-
-            /* Run remaining elements */
-            for(; x < input_width; ++x)
-            {
-                T element{};
-
-                if(is_log)
-                {
-                    element = (in_ptr[x] - max_val) * beta;
-                    sum += std::exp(element);
-                }
-                else
-                {
-                    element = std::exp((in_ptr[x] - max_val) * beta);
-                    sum += element;
-                }
-                tmp_ptr[x] = element;
-            }
-
-            if(!is_log)
-            {
-                sum_inversed = T(1) / sum;
-            }
-            else
-            {
-                sum = static_cast<T>(std::log(sum));
-            }
-        }
-
-        /* Normalize exponentials */
-        {
-            /* Loop over row and compute softmax */
-            int x = 0;
-            for(; x <= (input_width - vec_size); x += vec_size)
-            {
-                auto vec_in           = wrapper::vloadq(tmp_ptr + x);
-                auto normalized_value = wrapper::vdup_n(static_cast<T>(0), ExactTagType{});
-                if(is_log)
-                {
-                    normalized_value = wrapper::vsub(vec_in, wrapper::vdup_n(static_cast<T>(sum), ExactTagType{}));
-                }
-                else
-                {
-                    normalized_value = wrapper::vmul(vec_in, wrapper::vdup_n(static_cast<T>(sum_inversed), ExactTagType{}));
-                }
-                wrapper::vstore(out_ptr + x, normalized_value);
-            }
-            /* Run remaining elements */
-            for(; x < input_width; ++x)
-            {
-                if(is_log)
-                {
-                    out_ptr[x] = tmp_ptr[x] - sum;
-                }
-                else
-                {
-                    out_ptr[x] = tmp_ptr[x] * sum_inversed;
-                }
-            }
-        }
-    },
-    in_it, max_it, out_it);
-}
-
+                                  ITensor *out, const float beta, bool is_log, const Window &window);
 } // namespace cpu
 } // namespace arm_compute
 
diff --git a/src/cpu/kernels/softmax/generic/sve2/impl.cpp b/src/cpu/kernels/softmax/generic/sve2/impl.cpp
new file mode 100644
index 0000000000..9cdfe61446
--- /dev/null
+++ b/src/cpu/kernels/softmax/generic/sve2/impl.cpp
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2021-2022 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 "src/cpu/kernels/softmax/generic/sve2/impl.h"
+#include "arm_compute/core/Types.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <typename ScalarType>
+void sve2_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp,
+                                      ITensor *out, float beta, bool is_log, const Window &window)
+{
+    const int start_x     = in->info()->valid_region().anchor.x();
+    const int input_width = in->info()->valid_region().shape.x();
+
+    const float scale_beta     = -beta * in->info()->quantization_info().uniform().scale;
+    const auto  scale_beta_vec = svdup_n_f32(scale_beta);
+
+    Iterator   in_it(in, window);
+    Iterator   max_it(max, window);
+    Iterator   out_it(out, window);
+    const auto all_true_pg = wrapper::svptrue<ScalarType>();
+    using SVEType          = typename wrapper::traits::sve_vector<ScalarType>::type;
+
+    const int inc_1 = static_cast<int>(svcntw());
+    const int inc_2 = static_cast<int>(2 * svcntw());
+    const int inc_3 = static_cast<int>(3 * svcntw());
+
+    execute_window_loop(window, [&](const Coordinates &)
+    {
+        /* Get pointers */
+        const auto in_ptr  = reinterpret_cast<const ScalarType *>(in_it.ptr()) + start_x;
+        const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()) + start_x;
+        const auto tmp_ptr = reinterpret_cast<float *>(tmp);
+
+        float sum{};
+
+        /* Compute exponentials and sum */
+        {
+            /* Get max value */
+            const auto max_val = *reinterpret_cast<const ScalarType *>(max_it.ptr());
+            const auto vec_max = wrapper::svdup_n(max_val);
+
+            /* Init sum to zero */
+            auto vec_sum_0 = svdup_n_f32(0.f);
+            auto vec_sum_1 = svdup_n_f32(0.f);
+            auto vec_sum_2 = svdup_n_f32(0.f);
+            auto vec_sum_3 = svdup_n_f32(0.f);
+
+            /* Loop over row and compute exponentials and sum */
+            int      x    = 0;
+            svbool_t pg   = wrapper::svwhilelt<ScalarType>(x, input_width);
+            svbool_t pg_0 = svunpklo(svunpklo(pg));
+            svbool_t pg_1 = svunpkhi(svunpklo(pg));
+            svbool_t pg_2 = svunpklo(svunpkhi(pg));
+            svbool_t pg_3 = svunpkhi(svunpkhi(pg));
+            do
+            {
+                auto vec_elements = svld1(pg, in_ptr + x);
+                vec_elements      = svsub_z(pg, vec_max, vec_elements);
+
+                auto vec_elements_flt_0 = svcvt_f32_z(pg_0, svunpklo(svunpklo(vec_elements)));
+                auto vec_elements_flt_1 = svcvt_f32_z(pg_1, svunpkhi(svunpklo(vec_elements)));
+                auto vec_elements_flt_2 = svcvt_f32_z(pg_2, svunpklo(svunpkhi(vec_elements)));
+                auto vec_elements_flt_3 = svcvt_f32_z(pg_3, svunpkhi(svunpkhi(vec_elements)));
+
+                if(is_log)
+                {
+                    vec_elements_flt_0 = svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec);
+                    vec_elements_flt_1 = svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec);
+                    vec_elements_flt_2 = svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec);
+                    vec_elements_flt_3 = svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec);
+                    vec_sum_0          = svadd_f32_m(pg_0, vec_sum_0, svexp_f32_z(pg_0, vec_elements_flt_0));
+                    vec_sum_1          = svadd_f32_m(pg_1, vec_sum_1, svexp_f32_z(pg_1, vec_elements_flt_1));
+                    vec_sum_2          = svadd_f32_m(pg_2, vec_sum_2, svexp_f32_z(pg_2, vec_elements_flt_2));
+                    vec_sum_3          = svadd_f32_m(pg_3, vec_sum_3, svexp_f32_z(pg_3, vec_elements_flt_3));
+                }
+                else
+                {
+                    vec_elements_flt_0 = svexp_f32_z(pg_0, svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec));
+                    vec_elements_flt_1 = svexp_f32_z(pg_1, svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec));
+                    vec_elements_flt_2 = svexp_f32_z(pg_2, svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec));
+                    vec_elements_flt_3 = svexp_f32_z(pg_3, svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec));
+                    vec_sum_0          = svadd_f32_m(pg_0, vec_sum_0, vec_elements_flt_0);
+                    vec_sum_1          = svadd_f32_m(pg_1, vec_sum_1, vec_elements_flt_1);
+                    vec_sum_2          = svadd_f32_m(pg_2, vec_sum_2, vec_elements_flt_2);
+                    vec_sum_3          = svadd_f32_m(pg_3, vec_sum_3, vec_elements_flt_3);
+                }
+
+                svst1_f32(pg_0, tmp_ptr + x, vec_elements_flt_0);
+                svst1_f32(pg_1, tmp_ptr + x + inc_1, vec_elements_flt_1);
+                svst1_f32(pg_2, tmp_ptr + x + inc_2, vec_elements_flt_2);
+                svst1_f32(pg_3, tmp_ptr + x + inc_3, vec_elements_flt_3);
+
+                x += wrapper::svcnt<ScalarType>();
+                pg   = wrapper::svwhilelt<ScalarType>(x, input_width);
+                pg_0 = svunpklo(svunpklo(pg));
+                pg_1 = svunpkhi(svunpklo(pg));
+                pg_2 = svunpklo(svunpkhi(pg));
+                pg_3 = svunpkhi(svunpkhi(pg));
+            }
+            while(svptest_any(all_true_pg, pg));
+
+            /* Reduce sum */
+            const auto vec_sum = svadd_f32_z(all_true_pg, svadd_f32_z(all_true_pg, vec_sum_0, vec_sum_1), svadd_f32_z(all_true_pg, vec_sum_2, vec_sum_3));
+            sum                = svaddv_f32(all_true_pg, vec_sum);
+
+            /* Run remaining elements */
+            x = 0;
+            if(is_log)
+            {
+                sum = std::log(sum);
+            }
+            else
+            {
+                sum = 256.f / sum;
+            }
+        }
+
+        /* Normalize exponentials */
+        {
+            constexpr bool is_qasymm8_signed = std::is_same<ScalarType, qasymm8_signed_t>::value;
+            /* Loop over row and compute softmax */
+            int      x    = 0;
+            svbool_t pg   = wrapper::svwhilelt<ScalarType>(x, input_width);
+            svbool_t pg_0 = svunpklo(svunpklo(pg));
+            svbool_t pg_1 = svunpkhi(svunpklo(pg));
+            svbool_t pg_2 = svunpklo(svunpkhi(pg));
+            svbool_t pg_3 = svunpkhi(svunpkhi(pg));
+            do
+            {
+                auto vec_in_0 = svld1_f32(pg_0, tmp_ptr + x);
+                auto vec_in_1 = svld1_f32(pg_1, tmp_ptr + x + inc_1);
+                auto vec_in_2 = svld1_f32(pg_2, tmp_ptr + x + inc_2);
+                auto vec_in_3 = svld1_f32(pg_3, tmp_ptr + x + inc_3);
+
+                svfloat32_t res_0{};
+                svfloat32_t res_1{};
+                svfloat32_t res_2{};
+                svfloat32_t res_3{};
+
+                if(is_log)
+                {
+                    res_0 = svsub_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
+                    res_1 = svsub_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
+                    res_2 = svsub_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
+                    res_3 = svsub_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
+                }
+                else
+                {
+                    res_0 = svmul_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
+                    res_1 = svmul_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
+                    res_2 = svmul_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
+                    res_3 = svmul_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
+
+                    if(is_qasymm8_signed)
+                    {
+                        const auto offset_vec = svdup_n_f32(128.f);
+                        res_0                 = svsub_z(pg_0, vec_in_0, offset_vec);
+                        res_1                 = svsub_z(pg_1, vec_in_1, offset_vec);
+                        res_2                 = svsub_z(pg_2, vec_in_2, offset_vec);
+                        res_3                 = svsub_z(pg_3, vec_in_3, offset_vec);
+                    }
+                }
+
+                // Store value
+                const auto out = convert_float_to_int<SVEType>(res_0, res_1, res_2, res_3);
+                svst1(pg, out_ptr + x, out);
+                x += wrapper::svcnt<ScalarType>();
+                pg   = wrapper::svwhilelt<ScalarType>(x, input_width);
+                pg_0 = svunpklo(svunpklo(pg));
+                pg_1 = svunpkhi(svunpklo(pg));
+                pg_2 = svunpklo(svunpkhi(pg));
+                pg_3 = svunpkhi(svunpkhi(pg));
+            }
+            while(svptest_any(all_true_pg, pg));
+        }
+    },
+    in_it, max_it, out_it);
+}
+
+template void sve2_softmax_logits_1d_quantized<qasymm8_signed_t>(const ITensor *in, const ITensor *max, void *const tmp,
+                                                                 ITensor *out, float beta, bool is_log, const Window &window);
+template void sve2_softmax_logits_1d_quantized<qasymm8_t>(const ITensor *in, const ITensor *max, void *const tmp,
+                                                          ITensor *out, float beta, bool is_log, const Window &window);
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/cpu/kernels/softmax/generic/sve2/impl.h b/src/cpu/kernels/softmax/generic/sve2/impl.h
index 16dde2b115..abbcc15181 100644
--- a/src/cpu/kernels/softmax/generic/sve2/impl.h
+++ b/src/cpu/kernels/softmax/generic/sve2/impl.h
@@ -24,9 +24,7 @@
 #ifndef SRC_CORE_SVE2_KERNELS_SOFTMAX_IMPL_H
 #define SRC_CORE_SVE2_KERNELS_SOFTMAX_IMPL_H
 
-#if defined(ARM_COMPUTE_ENABLE_SVE2)
-#include "arm_compute/core/Types.h"
-#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "arm_compute/core/Helpers.h"
 
 namespace arm_compute
 {
@@ -34,177 +32,7 @@ namespace cpu
 {
 template <typename ScalarType>
 void sve2_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp,
-                                      ITensor *out, float beta, bool is_log, const Window &window)
-{
-    const int start_x     = in->info()->valid_region().anchor.x();
-    const int input_width = in->info()->valid_region().shape.x();
-
-    const float scale_beta     = -beta * in->info()->quantization_info().uniform().scale;
-    const auto  scale_beta_vec = svdup_n_f32(scale_beta);
-
-    Iterator   in_it(in, window);
-    Iterator   max_it(max, window);
-    Iterator   out_it(out, window);
-    const auto all_true_pg = wrapper::svptrue<ScalarType>();
-    using SVEType          = typename wrapper::traits::sve_vector<ScalarType>::type;
-
-    const int inc_1 = static_cast<int>(svcntw());
-    const int inc_2 = static_cast<int>(2 * svcntw());
-    const int inc_3 = static_cast<int>(3 * svcntw());
-
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        /* Get pointers */
-        const auto in_ptr  = reinterpret_cast<const ScalarType *>(in_it.ptr()) + start_x;
-        const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()) + start_x;
-        const auto tmp_ptr = reinterpret_cast<float *>(tmp);
-
-        float sum{};
-
-        /* Compute exponentials and sum */
-        {
-            /* Get max value */
-            const auto max_val = *reinterpret_cast<const ScalarType *>(max_it.ptr());
-            const auto vec_max = wrapper::svdup_n(max_val);
-
-            /* Init sum to zero */
-            auto vec_sum_0 = svdup_n_f32(0.f);
-            auto vec_sum_1 = svdup_n_f32(0.f);
-            auto vec_sum_2 = svdup_n_f32(0.f);
-            auto vec_sum_3 = svdup_n_f32(0.f);
-
-            /* Loop over row and compute exponentials and sum */
-            int      x    = 0;
-            svbool_t pg   = wrapper::svwhilelt<ScalarType>(x, input_width);
-            svbool_t pg_0 = svunpklo(svunpklo(pg));
-            svbool_t pg_1 = svunpkhi(svunpklo(pg));
-            svbool_t pg_2 = svunpklo(svunpkhi(pg));
-            svbool_t pg_3 = svunpkhi(svunpkhi(pg));
-            do
-            {
-                auto vec_elements = svld1(pg, in_ptr + x);
-                vec_elements      = svsub_z(pg, vec_max, vec_elements);
-
-                auto vec_elements_flt_0 = svcvt_f32_z(pg_0, svunpklo(svunpklo(vec_elements)));
-                auto vec_elements_flt_1 = svcvt_f32_z(pg_1, svunpkhi(svunpklo(vec_elements)));
-                auto vec_elements_flt_2 = svcvt_f32_z(pg_2, svunpklo(svunpkhi(vec_elements)));
-                auto vec_elements_flt_3 = svcvt_f32_z(pg_3, svunpkhi(svunpkhi(vec_elements)));
-
-                if(is_log)
-                {
-                    vec_elements_flt_0 = svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec);
-                    vec_elements_flt_1 = svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec);
-                    vec_elements_flt_2 = svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec);
-                    vec_elements_flt_3 = svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec);
-                    vec_sum_0          = svadd_f32_m(pg_0, vec_sum_0, svexp_f32_z(pg_0, vec_elements_flt_0));
-                    vec_sum_1          = svadd_f32_m(pg_1, vec_sum_1, svexp_f32_z(pg_1, vec_elements_flt_1));
-                    vec_sum_2          = svadd_f32_m(pg_2, vec_sum_2, svexp_f32_z(pg_2, vec_elements_flt_2));
-                    vec_sum_3          = svadd_f32_m(pg_3, vec_sum_3, svexp_f32_z(pg_3, vec_elements_flt_3));
-                }
-                else
-                {
-                    vec_elements_flt_0 = svexp_f32_z(pg_0, svmul_f32_z(pg_0, vec_elements_flt_0, scale_beta_vec));
-                    vec_elements_flt_1 = svexp_f32_z(pg_1, svmul_f32_z(pg_1, vec_elements_flt_1, scale_beta_vec));
-                    vec_elements_flt_2 = svexp_f32_z(pg_2, svmul_f32_z(pg_2, vec_elements_flt_2, scale_beta_vec));
-                    vec_elements_flt_3 = svexp_f32_z(pg_3, svmul_f32_z(pg_3, vec_elements_flt_3, scale_beta_vec));
-                    vec_sum_0          = svadd_f32_m(pg_0, vec_sum_0, vec_elements_flt_0);
-                    vec_sum_1          = svadd_f32_m(pg_1, vec_sum_1, vec_elements_flt_1);
-                    vec_sum_2          = svadd_f32_m(pg_2, vec_sum_2, vec_elements_flt_2);
-                    vec_sum_3          = svadd_f32_m(pg_3, vec_sum_3, vec_elements_flt_3);
-                }
-
-                svst1_f32(pg_0, tmp_ptr + x, vec_elements_flt_0);
-                svst1_f32(pg_1, tmp_ptr + x + inc_1, vec_elements_flt_1);
-                svst1_f32(pg_2, tmp_ptr + x + inc_2, vec_elements_flt_2);
-                svst1_f32(pg_3, tmp_ptr + x + inc_3, vec_elements_flt_3);
-
-                x += wrapper::svcnt<ScalarType>();
-                pg   = wrapper::svwhilelt<ScalarType>(x, input_width);
-                pg_0 = svunpklo(svunpklo(pg));
-                pg_1 = svunpkhi(svunpklo(pg));
-                pg_2 = svunpklo(svunpkhi(pg));
-                pg_3 = svunpkhi(svunpkhi(pg));
-            }
-            while(svptest_any(all_true_pg, pg));
-
-            /* Reduce sum */
-            const auto vec_sum = svadd_f32_z(all_true_pg, svadd_f32_z(all_true_pg, vec_sum_0, vec_sum_1), svadd_f32_z(all_true_pg, vec_sum_2, vec_sum_3));
-            sum                = svaddv_f32(all_true_pg, vec_sum);
-
-            /* Run remaining elements */
-            x = 0;
-            if(is_log)
-            {
-                sum = std::log(sum);
-            }
-            else
-            {
-                sum = 256.f / sum;
-            }
-        }
-
-        /* Normalize exponentials */
-        {
-            constexpr bool is_qasymm8_signed = std::is_same<ScalarType, qasymm8_signed_t>::value;
-            /* Loop over row and compute softmax */
-            int      x    = 0;
-            svbool_t pg   = wrapper::svwhilelt<ScalarType>(x, input_width);
-            svbool_t pg_0 = svunpklo(svunpklo(pg));
-            svbool_t pg_1 = svunpkhi(svunpklo(pg));
-            svbool_t pg_2 = svunpklo(svunpkhi(pg));
-            svbool_t pg_3 = svunpkhi(svunpkhi(pg));
-            do
-            {
-                auto vec_in_0 = svld1_f32(pg_0, tmp_ptr + x);
-                auto vec_in_1 = svld1_f32(pg_1, tmp_ptr + x + inc_1);
-                auto vec_in_2 = svld1_f32(pg_2, tmp_ptr + x + inc_2);
-                auto vec_in_3 = svld1_f32(pg_3, tmp_ptr + x + inc_3);
-
-                svfloat32_t res_0{};
-                svfloat32_t res_1{};
-                svfloat32_t res_2{};
-                svfloat32_t res_3{};
-
-                if(is_log)
-                {
-                    res_0 = svsub_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
-                    res_1 = svsub_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
-                    res_2 = svsub_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
-                    res_3 = svsub_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
-                }
-                else
-                {
-                    res_0 = svmul_f32_z(pg_0, vec_in_0, svdup_n_f32(sum));
-                    res_1 = svmul_f32_z(pg_1, vec_in_1, svdup_n_f32(sum));
-                    res_2 = svmul_f32_z(pg_2, vec_in_2, svdup_n_f32(sum));
-                    res_3 = svmul_f32_z(pg_3, vec_in_3, svdup_n_f32(sum));
-
-                    if(is_qasymm8_signed)
-                    {
-                        const auto offset_vec = svdup_n_f32(128.f);
-                        res_0                 = svsub_z(pg_0, vec_in_0, offset_vec);
-                        res_1                 = svsub_z(pg_1, vec_in_1, offset_vec);
-                        res_2                 = svsub_z(pg_2, vec_in_2, offset_vec);
-                        res_3                 = svsub_z(pg_3, vec_in_3, offset_vec);
-                    }
-                }
-
-                // Store value
-                const auto out = convert_float_to_int<SVEType>(res_0, res_1, res_2, res_3);
-                svst1(pg, out_ptr + x, out);
-                x += wrapper::svcnt<ScalarType>();
-                pg   = wrapper::svwhilelt<ScalarType>(x, input_width);
-                pg_0 = svunpklo(svunpklo(pg));
-                pg_1 = svunpkhi(svunpklo(pg));
-                pg_2 = svunpklo(svunpkhi(pg));
-                pg_3 = svunpkhi(svunpkhi(pg));
-            }
-            while(svptest_any(all_true_pg, pg));
-        }
-    },
-    in_it, max_it, out_it);
-}
+                                      ITensor *out, float beta, bool is_log, const Window &window);
 } // namespace cpu
 } // namespace arm_compute
-#endif /* defined(ARM_COMPUTE_ENABLE_SVE2) */
 #endif /* SRC_CORE_SVE2_KERNELS_SOFTMAX_IMPL_H */