Add SVE for Softmax

Implements COMPMID-3875

Change-Id: I38991eed3f4966db125862af066bfedff5994a25
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4854
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 1023 insertions, 443 deletions

diff --git a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp
index 97797cefde..fe09f1ec59 100644
--- a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp
+++ b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -27,61 +27,169 @@
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
-#include "src/core/AccessWindowStatic.h"
 #include "src/core/CPP/Validate.h"
-#include "src/core/NEON/NEFixedPoint.h"
-#include "src/core/NEON/NEMath.h"
-#include "src/core/NEON/wrapper/wrapper.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
-#include "support/SaturateCast.h"
 
-#include <algorithm>
-#include <arm_neon.h>
-#include <cfloat>
-#include <functional>
+#include "src/core/NEON/kernels/softmax/impl/NEON/list.h"
+#include "src/core/NEON/kernels/softmax/impl/SVE/list.h"
+#include "src/core/common/Registrars.h"
 
 namespace arm_compute
 {
-template <typename float_vec_type, typename int_vec_type>
-int_vec_type convert_float_to_int(const float_vec_type &in);
+namespace
+{
+struct SoftmaxSelectorData
+{
+    DataType dt;
+};
+using SoftmaxSelectorPtr          = std::add_pointer<bool(const SoftmaxSelectorData &data)>::type;
+using SoftmaxLogits1DMaxKernelPtr = std::add_pointer<void(const ITensor *, ITensor *, const Window &)>::type;
+using SoftmaxLogits1DKernelPtr    = std::add_pointer<void(const ITensor *, const ITensor *, void *const, ITensor *, float, bool, const Window &)>::type;
 
-template <typename float_vec_type, typename int_vec_type>
-float_vec_type convert_int_to_float(const int_vec_type &in);
+struct SoftmaxLogits1DKernel
+{
+    const char              *name;
+    const SoftmaxSelectorPtr is_selected;
+    SoftmaxLogits1DKernelPtr ukernel;
+};
 
-template <>
-uint8x16_t convert_float_to_int<float32x4x4_t, uint8x16_t>(const float32x4x4_t &in)
+struct SoftmaxLogits1DMaxKernel
 {
-    uint8x16_t out;
-    convert_float32x4x4_to_uint8x16(in, out);
-    return out;
-}
+    const char                 *name;
+    const SoftmaxSelectorPtr    is_selected;
+    SoftmaxLogits1DMaxKernelPtr ukernel;
+};
 
-template <>
-int8x16_t convert_float_to_int<float32x4x4_t, int8x16_t>(const float32x4x4_t &in)
+static const SoftmaxLogits1DKernel available_logits_1d_kernels[] =
 {
-    int8x16_t out;
-    convert_float32x4x4_to_int8x16(in, out);
-    return out;
-}
+#if defined(__ARM_FEATURE_SVE)
+    {
+        "sve_softmax_logits_1d_float",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::F32); },
+        REGISTER_FP32_SVE(arm_compute::cpu::sve_softmax_logits_1d_float<float>)
+    },
+    {
+        "sve_softmax_logits_1d_float",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::F16); },
+        REGISTER_FP16_SVE(arm_compute::cpu::sve_softmax_logits_1d_float<float16_t>)
+    },
+#else /* !defined(__ARM_FEATURE_SVE) */
+    {
+        "neon_softmax_logits_1d_float",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::F32); },
+        REGISTER_FP32_NEON(arm_compute::cpu::neon_softmax_logits_1d_float<float>)
+    },
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    {
+        "neon_softmax_logits_1d_float",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::F16); },
+        REGISTER_FP16_NEON(arm_compute::cpu::neon_softmax_logits_1d_float<float16_t>)
+    },
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
+#endif /* defined(__ARM_FEATURE_SVE) */
 
-template <>
-float32x4x4_t convert_int_to_float<float32x4x4_t, uint8x16_t>(const uint8x16_t &in)
+#if defined(__ARM_FEATURE_SVE2)
+    {
+        "sve_softmax_logits_1d_quantized",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::QASYMM8); },
+        REGISTER_QASYMM8_SVE(arm_compute::cpu::sve_softmax_logits_1d_quantized<qasymm8_t>)
+    },
+    {
+        "sve_softmax_logits_1d_quantized",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::QASYMM8_SIGNED); },
+        REGISTER_QASYMM8_SIGNED_SVE(arm_compute::cpu::sve_softmax_logits_1d_quantized<qasymm8_signed_t>)
+    },
+#else  /* !defined(__ARM_FEATURE_SVE2) */
+    {
+        "neon_softmax_logits_1d_quantized",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::QASYMM8); },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_softmax_logits_1d_quantized<qasymm8_t>)
+    },
+    {
+        "neon_softmax_logits_1d_quantized",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::QASYMM8_SIGNED); },
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::neon_softmax_logits_1d_quantized<qasymm8_signed_t>)
+    },
+#endif /* defined(__ARM_FEATURE_SVE2) */
+
+};
+
+static const SoftmaxLogits1DMaxKernel available_logits_1d_max_kernels[] =
 {
-    return convert_uint8x16_to_float32x4x4(in);
-}
+#if defined(__ARM_FEATURE_SVE)
+    {
+        "sve_logits_1d_max",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::F32); },
+        REGISTER_FP32_SVE(arm_compute::cpu::sve_logits_1d_max<float>)
+    },
+    {
+        "sve_logits_1d_max",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::F16); },
+        REGISTER_FP16_SVE(arm_compute::cpu::sve_logits_1d_max<float16_t>)
+    },
+    {
+        "sve_logits_1d_max",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::QASYMM8); },
+        REGISTER_QASYMM8_SVE(arm_compute::cpu::sve_logits_1d_max<qasymm8_t>)
+    },
+    {
+        "sve_logits_1d_max",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::QASYMM8_SIGNED); },
+        REGISTER_QASYMM8_SIGNED_SVE(arm_compute::cpu::sve_logits_1d_max<qasymm8_signed_t>)
+    },
+#else /* !defined(__ARM_FEATURE_SVE) */
+    {
+        "neon_logits_1d_max",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::F32); },
+        REGISTER_FP32_NEON(arm_compute::cpu::neon_logits_1d_max<float>)
+    },
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    {
+        "neon_logits_1d_max",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::F16); },
+        REGISTER_FP16_NEON(arm_compute::cpu::neon_logits_1d_max<float16_t>)
+    },
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
+    {
+        "neon_logits_1d_max",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::QASYMM8); },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_logits_1d_max<qasymm8_t>)
+    },
+    {
+        "neon_logits_1d_max",
+        [](const SoftmaxSelectorData & data) { return (data.dt == DataType::QASYMM8_SIGNED); },
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::neon_logits_1d_max<qasymm8_signed_t>)
+    },
+#endif /* defined(__ARM_FEATURE_SVE) */
+};
 
-template <>
-float32x4x4_t convert_int_to_float<float32x4x4_t, int8x16_t>(const int8x16_t &in)
+const SoftmaxLogits1DKernel *get_implementation_logits(const SoftmaxSelectorData &data)
 {
-    return convert_int8x16_to_float32x4x4(in);
+    for(const auto &uk : available_logits_1d_kernels)
+    {
+        if(uk.is_selected({ data.dt }))
+        {
+            return &uk;
+        }
+    }
+    return nullptr;
 }
 
-namespace
+const SoftmaxLogits1DMaxKernel *get_implementation_logits_max(const SoftmaxSelectorData &data)
 {
+    for(const auto &uk : available_logits_1d_max_kernels)
+    {
+        if(uk.is_selected({ data.dt }))
+        {
+            return &uk;
+        }
+    }
+    return nullptr;
+}
+
 Status validate_arguments_logits_1d_max(const ITensorInfo &input, const ITensorInfo &output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input);
@@ -98,59 +206,10 @@ Status validate_arguments_logits_1d_max(const ITensorInfo &input, const ITensorI
     return Status{};
 }
 
-template <typename T>
-void logits_1d_max(const ITensor &in, ITensor &out, const Window &window)
-{
-    /** NEON 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);
-}
 } // namespace
 
 NELogits1DMaxKernel::NELogits1DMaxKernel()
-    : _func(nullptr), _border_size()
+    : _border_size()
 {
 }
 
@@ -177,26 +236,6 @@ void NELogits1DMaxKernel::configure(const ITensor *input, ITensor *output)
     coord.set_num_dimensions(output->info()->num_dimensions());
     output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
 
-    switch(input->info()->data_type())
-    {
-        case DataType::QASYMM8:
-            _func = &logits_1d_max<qasymm8_t>;
-            break;
-        case DataType::QASYMM8_SIGNED:
-            _func = &logits_1d_max<qasymm8_signed_t>;
-            break;
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        case DataType::F16:
-            _func = &logits_1d_max<float16_t>;
-            break;
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-        case DataType::F32:
-            _func = &logits_1d_max<float>;
-            break;
-        default:
-            ARM_COMPUTE_ERROR("Unsupported data type.");
-    }
-
     _input  = input;
     _output = output;
 
@@ -222,9 +261,9 @@ void NELogits1DMaxKernel::run(const Window &window, const ThreadInfo &info)
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
-    ARM_COMPUTE_ERROR_ON(_func == nullptr);
 
-    (*_func)(*_input, *_output, window);
+    const auto *uk = get_implementation_logits_max(SoftmaxSelectorData{ _input->info()->data_type() });
+    uk->ukernel(_input, _output, window);
 }
 
 namespace
@@ -265,308 +304,11 @@ Status validate_arguments_logits_softmax(const ITensorInfo &input, const ITensor
 
     return Status{};
 }
-template <typename T, bool is_log>
-void logits_1d_softmax_qasymm8(const ITensor &in, const ITensor &max, void *const tmp, ITensor &out, const float beta, 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 <typename T, bool is_log = false>
-void logits_1d_softmax_float(const ITensor &in, const ITensor &max, void *const tmp,
-                             ITensor &out, const float beta, 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);
-
-    /** NEON 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);
-}
 } // namespace
 
 template <bool IS_LOG>
 NELogits1DSoftmaxKernel<IS_LOG>::NELogits1DSoftmaxKernel()
-    : _func(nullptr), _input(nullptr), _max(nullptr), _output(nullptr), _beta(1.0f), _tmp(nullptr)
+    : _input(nullptr), _max(nullptr), _output(nullptr), _beta(1.0f), _tmp(nullptr)
 {
 }
 
@@ -595,27 +337,6 @@ void NELogits1DSoftmaxKernel<IS_LOG>::configure(const ITensor *input, const ITen
     coord.set_num_dimensions(output->info()->num_dimensions());
     output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
 
-    switch(input->info()->data_type())
-    {
-        case DataType::QASYMM8:
-            _func = &logits_1d_softmax_qasymm8<qasymm8_t, IS_LOG>;
-            break;
-        case DataType::QASYMM8_SIGNED:
-            _func = &logits_1d_softmax_qasymm8<qasymm8_signed_t, IS_LOG>;
-            break;
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        case DataType::F16:
-            _func = &logits_1d_softmax_float<float16_t, IS_LOG>;
-            break;
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-        case DataType::F32:
-            _func = &logits_1d_softmax_float<float, IS_LOG>;
-            break;
-        default:
-            ARM_COMPUTE_ERROR("Unsupported data type.");
-            break;
-    }
-
     _input  = input;
     _max    = max;
     _output = output;
@@ -649,7 +370,8 @@ void NELogits1DSoftmaxKernel<IS_LOG>::run(const Window &window, const ThreadInfo
 
     void *tmp_for_thread = _tmp->buffer() + (info.thread_id * tmp_size_for_thread);
 
-    (*_func)(*_input, *_max, tmp_for_thread, *_output, _beta, window);
+    const auto *uk = get_implementation_logits(SoftmaxSelectorData{ _input->info()->data_type() });
+    uk->ukernel(_input, _max, tmp_for_thread, _output, _beta, IS_LOG, window);
 }
 
 template class NELogits1DSoftmaxKernel<true>;
diff --git a/src/core/NEON/kernels/NESoftmaxLayerKernel.h b/src/core/NEON/kernels/NESoftmaxLayerKernel.h
index adc2e57258..70e2417fc2 100644
--- a/src/core/NEON/kernels/NESoftmaxLayerKernel.h
+++ b/src/core/NEON/kernels/NESoftmaxLayerKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -71,11 +71,7 @@ public:
     BorderSize border_size() const override;
 
 private:
-    using Logits1DMaxFunction = void(const ITensor &in, ITensor &out, const Window &window);
-
-private:
-    Logits1DMaxFunction *_func;
-    BorderSize           _border_size;
+    BorderSize _border_size;
 };
 
 /** Interface for softmax computation for QASYMM8 with pre-computed max. */
@@ -135,15 +131,11 @@ public:
     void run(const Window &window, const ThreadInfo &info) override;
 
 private:
-    using LogitsSoftmaxFunction = void(const ITensor &in, const ITensor &max, void *const tmp, ITensor &out, const float beta,
-                                       const Window &window);
-
-    LogitsSoftmaxFunction *_func;
-    const ITensor         *_input;
-    const ITensor         *_max;
-    ITensor               *_output;
-    float                  _beta;
-    ITensor               *_tmp; //Temporary. Used internally
+    const ITensor *_input;
+    const ITensor *_max;
+    ITensor       *_output;
+    float          _beta;
+    ITensor       *_tmp; //Temporary. Used internally
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_NESOFTMAXLAYERKERNEL_H */
diff --git a/src/core/NEON/kernels/softmax/impl/NEON/list.h b/src/core/NEON/kernels/softmax/impl/NEON/list.h
new file mode 100644
index 0000000000..a8f781f439
--- /dev/null
+++ b/src/core/NEON/kernels/softmax/impl/NEON/list.h
@@ -0,0 +1,425 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_SOFTMAX_LIST_H
+#define SRC_CORE_NEON_KERNELS_SOFTMAX_LIST_H
+
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "support/SaturateCast.h"
+#include "src/core/NEON/NEFixedPoint.h"
+#include "src/core/NEON/NEMath.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace
+{
+template <typename float_vec_type, typename int_vec_type>
+int_vec_type convert_float_to_int(const float_vec_type &in);
+
+template <typename float_vec_type, typename int_vec_type>
+float_vec_type convert_int_to_float(const int_vec_type &in);
+
+template <>
+uint8x16_t convert_float_to_int<float32x4x4_t, uint8x16_t>(const float32x4x4_t &in)
+{
+    uint8x16_t out;
+    convert_float32x4x4_to_uint8x16(in, out);
+    return out;
+}
+
+template <>
+int8x16_t convert_float_to_int<float32x4x4_t, int8x16_t>(const float32x4x4_t &in)
+{
+    int8x16_t out;
+    convert_float32x4x4_to_int8x16(in, out);
+    return out;
+}
+
+template <>
+float32x4x4_t convert_int_to_float<float32x4x4_t, uint8x16_t>(const uint8x16_t &in)
+{
+    return convert_uint8x16_to_float32x4x4(in);
+}
+
+template <>
+float32x4x4_t convert_int_to_float<float32x4x4_t, int8x16_t>(const int8x16_t &in)
+{
+    return convert_int8x16_to_float32x4x4(in);
+}
+} // namespace
+
+template <typename T>
+void neon_logits_1d_max(const ITensor *in, ITensor *out, const Window &window)
+{
+    /** NEON 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);
+}
+
+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 <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);
+
+    /** NEON 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);
+}
+
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* SRC_CORE_NEON_KERNELS_SOFTMAX_LIST_H */
diff --git a/src/core/NEON/kernels/softmax/impl/SVE/list.h b/src/core/NEON/kernels/softmax/impl/SVE/list.h
new file mode 100644
index 0000000000..4f3c07d2a2
--- /dev/null
+++ b/src/core/NEON/kernels/softmax/impl/SVE/list.h
@@ -0,0 +1,427 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef SRC_CORE_SVE_KERNELS_SOFTMAX_LIST_H
+#define SRC_CORE_SVE_KERNELS_SOFTMAX_LIST_H
+
+#if defined(__ARM_FEATURE_SVE)
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/SVEMath.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace
+{
+template <typename int_vec_type>
+int_vec_type convert_float_to_int(const svfloat32_t &in_0, const svfloat32_t &in_1, const svfloat32_t &in_2, const svfloat32_t &in_3);
+
+template <>
+svuint8_t convert_float_to_int<svuint8_t>(const svfloat32_t &in_0, const svfloat32_t &in_1, const svfloat32_t &in_2, const svfloat32_t &in_3)
+{
+    svuint8_t  out;
+    const auto all_true_pg = svptrue_b32();
+    auto       tmp_0       = svcvt_u32_f32_z(all_true_pg, in_0);
+    auto       tmp_1       = svcvt_u32_f32_z(all_true_pg, in_1);
+    auto       tmp_2       = svcvt_u32_f32_z(all_true_pg, in_2);
+    auto       tmp_3       = svcvt_u32_f32_z(all_true_pg, in_3);
+
+    auto tmp_16_0 = svqxtnt_u32(svqxtnb_u32(tmp_0), tmp_1);
+    auto tmp_16_1 = svqxtnt_u32(svqxtnb_u32(tmp_2), tmp_3);
+
+    auto tmp_16_uzp_0 = svuzp1(tmp_16_0, tmp_16_0);
+    auto tmp_16_uzp_1 = svuzp2(tmp_16_0, tmp_16_0);
+    auto tmp_16_uzp_2 = svuzp1(tmp_16_1, tmp_16_1);
+    auto tmp_16_uzp_3 = svuzp2(tmp_16_1, tmp_16_1);
+
+    auto pg = svwhilelt_b16_s32(0, svcnth() / 2);
+
+    tmp_16_0 = svsplice(pg, tmp_16_uzp_0, tmp_16_uzp_1);
+    tmp_16_1 = svsplice(pg, tmp_16_uzp_2, tmp_16_uzp_3);
+
+    out = svqxtnt_u16(svqxtnb_u16(tmp_16_0), tmp_16_1);
+
+    auto out_uzp_0 = svuzp1(out, out);
+    auto out_uzp_1 = svuzp2(out, out);
+
+    pg  = svwhilelt_b8_s32(0, svcntb() / 2);
+    out = svsplice(pg, out_uzp_0, out_uzp_1);
+
+    return out;
+}
+
+template <>
+svint8_t convert_float_to_int<svint8_t>(const svfloat32_t &in_0, const svfloat32_t &in_1, const svfloat32_t &in_2, const svfloat32_t &in_3)
+{
+    svint8_t   out;
+    const auto all_true_pg = svptrue_b32();
+    auto       tmp_0       = svcvt_s32_f32_z(all_true_pg, in_0);
+    auto       tmp_1       = svcvt_s32_f32_z(all_true_pg, in_1);
+    auto       tmp_2       = svcvt_s32_f32_z(all_true_pg, in_2);
+    auto       tmp_3       = svcvt_s32_f32_z(all_true_pg, in_3);
+
+    auto tmp_16_0 = svqxtnt_s32(svqxtnb_s32(tmp_0), tmp_1);
+    auto tmp_16_1 = svqxtnt_s32(svqxtnb_s32(tmp_2), tmp_3);
+
+    auto tmp_16_uzp_0 = svuzp1(tmp_16_0, tmp_16_0);
+    auto tmp_16_uzp_1 = svuzp2(tmp_16_0, tmp_16_0);
+    auto tmp_16_uzp_2 = svuzp1(tmp_16_1, tmp_16_1);
+    auto tmp_16_uzp_3 = svuzp2(tmp_16_1, tmp_16_1);
+
+    auto pg = svwhilelt_b16_s32(0, svcnth() / 2);
+
+    tmp_16_0 = svsplice(pg, tmp_16_uzp_0, tmp_16_uzp_1);
+    tmp_16_1 = svsplice(pg, tmp_16_uzp_2, tmp_16_uzp_3);
+
+    out = svqxtnt_s16(svqxtnb_s16(tmp_16_0), tmp_16_1);
+
+    auto out_uzp_0 = svuzp1(out, out);
+    auto out_uzp_1 = svuzp2(out, out);
+
+    pg  = svwhilelt_b8_s32(0, svcntb() / 2);
+    out = svsplice(pg, out_uzp_0, out_uzp_1);
+
+    return out;
+}
+} // namespace
+
+template <typename ScalarType>
+void sve_logits_1d_max(const ITensor *in, ITensor *out, const Window &window)
+{
+    const auto all_true_pg    = wrapper::svptrue<ScalarType>();
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Window win{ window };
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator input(in, win);
+    Iterator output(out, win);
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        // Get pointers
+        const auto in_ptr  = reinterpret_cast<const ScalarType *>(input.ptr());
+        const auto out_ptr = reinterpret_cast<ScalarType *>(output.ptr());
+
+        // Init max value
+        auto vec_max = wrapper::svdup_n(support::cpp11::lowest<ScalarType>());
+
+        int      x  = window_start_x;
+        svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
+        do
+        {
+            const auto current_value = svld1(pg, in_ptr + x);
+            vec_max                  = svmax_m(pg, vec_max, current_value);
+
+            x += wrapper::svcnt<ScalarType>();
+            pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
+        }
+        while(svptest_any(all_true_pg, pg));
+
+        auto max_val = svmaxv(all_true_pg, vec_max);
+
+        *out_ptr = max_val;
+    },
+    input, output);
+}
+
+#if defined(__ARM_FEATURE_SVE2)
+template <typename ScalarType>
+void sve_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);
+}
+#endif /* defined(__ARM_FEATURE_SVE2) */
+
+template <typename ScalarType>
+void sve_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);
+
+    const auto all_true_pg = wrapper::svptrue<ScalarType>();
+
+    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<ScalarType *>(tmp);
+
+        ScalarType sum{ 0 };
+
+        /* 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 = wrapper::svdup_n(static_cast<ScalarType>(0));
+
+            /* Loop over row and compute exponentials and sum */
+            int      x  = 0;
+            svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
+            do
+            {
+                auto vec_elements = svld1(pg, in_ptr + x);
+                vec_elements      = svsub_z(pg, vec_elements, vec_max);
+                if(is_log)
+                {
+                    vec_elements = svmul_z(pg, vec_elements, wrapper::svdup_n(static_cast<ScalarType>(beta)));
+                    vec_sum      = svadd_m(pg, vec_sum, wrapper::svexp_z(pg, vec_elements));
+                }
+                else
+                {
+                    vec_elements = wrapper::svexp_z(pg, svmul_z(pg, vec_elements, wrapper::svdup_n(static_cast<ScalarType>(beta))));
+                    vec_sum      = svadd_m(pg, vec_sum, vec_elements);
+                }
+                svst1(pg, tmp_ptr + x, vec_elements);
+
+                x += wrapper::svcnt<ScalarType>();
+                pg = wrapper::svwhilelt<ScalarType>(x, input_width);
+            }
+            while(svptest_any(all_true_pg, pg));
+
+            /* Reduce sum */
+            sum = svaddv(all_true_pg, vec_sum);
+
+            if(is_log)
+            {
+                sum = static_cast<ScalarType>(std::log(sum));
+            }
+            else
+            {
+                sum = ScalarType(1) / sum;
+            }
+        }
+
+        /* Normalize exponentials */
+        {
+            /* Loop over row and compute softmax */
+            int      x  = 0;
+            svbool_t pg = wrapper::svwhilelt<ScalarType>(x, input_width);
+            do
+            {
+                auto vec_in           = svld1(pg, tmp_ptr + x);
+                auto normalized_value = wrapper::svdup_n(static_cast<ScalarType>(0));
+                if(is_log)
+                {
+                    normalized_value = svsub_z(pg, vec_in, wrapper::svdup_n(static_cast<ScalarType>(sum)));
+                }
+                else
+                {
+                    normalized_value = svmul_z(pg, vec_in, wrapper::svdup_n(static_cast<ScalarType>(sum)));
+                }
+                svst1(pg, out_ptr + x, normalized_value);
+
+                x += wrapper::svcnt<ScalarType>();
+                pg = wrapper::svwhilelt<ScalarType>(x, input_width);
+            }
+            while(svptest_any(all_true_pg, pg));
+        }
+    },
+    in_it, max_it, out_it);
+}
+
+} // namespace cpu
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_SVE) */
+
+#endif /* SRC_CORE_SVE_KERNELS_SOFTMAX_LIST_H */
diff --git a/src/core/NEON/wrapper/traits.h b/src/core/NEON/wrapper/traits.h
index eafbeef372..0bb55afe02 100644
--- a/src/core/NEON/wrapper/traits.h
+++ b/src/core/NEON/wrapper/traits.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,6 +26,10 @@
 
 #include <arm_neon.h>
 
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+#endif /* defined(__ARM_FEATURE_SVE) */
+
 namespace arm_compute
 {
 namespace wrapper
@@ -110,6 +114,16 @@ template <> struct neon_bitvector<float_t, BitWidth::W128>{ using type = float32
 template <> struct neon_bitvector<float16_t, BitWidth::W64>{ using type = float16x4_t; using tag_type = vector_64_tag; };
 template <> struct neon_bitvector<float16_t, BitWidth::W128>{ using type = float16x8_t; using tag_type = vector_128_tag; };
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+
+#if defined(__ARM_FEATURE_SVE)
+/** Create the appropriate SVE vector given its type */
+template <typename T> struct sve_vector;
+
+template <> struct sve_vector<uint8_t>{ using scalar_type = uint8_t; using type = svuint8_t; };
+template <> struct sve_vector<int8_t>{ using scalar_type = int8_t; using type = svint8_t; };
+#endif /* defined(__ARM_FEATURE_SVE) */
+
 #endif /* DOXYGEN_SKIP_THIS */
 
 /**  Helper type template to get the type of a neon vector */