Make CpuActivation stateless

- Rename NEActivationLayer to CpuActivation
- Add member function to generate execution window

Partially Resolves: COMPMID-3992

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I4e1ae15cf456b860d3080b2fedc4dbcce7d1bb79
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4791
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

5 files changed, 1089 insertions, 0 deletions

diff --git a/src/core/cpu/kernels/activation/NEON/fp16.cpp b/src/core/cpu/kernels/activation/NEON/fp16.cpp
new file mode 100644
index 0000000000..27ae2830cc
--- /dev/null
+++ b/src/core/cpu/kernels/activation/NEON/fp16.cpp
@@ -0,0 +1,215 @@
+/*
+ * Copyright (c) 2020-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "src/core/NEON/NEMath.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Validate.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace
+{
+#ifndef __aarch64__
+inline float16x8_t mask_float_vector(const float16x8_t &in, const uint16x8_t &mask)
+{
+    auto int_in = vreinterpretq_u16_f16(in);
+    return vreinterpretq_f16_u16(wrapper::vand(int_in, mask));
+}
+#endif /* __arch64__ */
+} // namespace
+
+void fp16_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+{
+    /** NEON vector tag type. */
+    using ExactTagType                                = typename wrapper::traits::neon_bitvector_tag_t<float16_t, wrapper::traits::BitWidth::W128>;
+    const ActivationLayerInfo::ActivationFunction act = act_info.activation();
+
+    constexpr int window_step_x  = 8;
+    const auto    window_start_x = static_cast<int>(window.x().start());
+    const auto    window_end_x   = static_cast<int>(window.x().end());
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    // In case of non-aarch64, a small delta value is added to the input
+    // to prevent NAN values caused by zeros in inputs to SQRT.
+    // In case of aarh64, we call vsqrt directly, so we don't use delta.
+#ifndef __aarch64__
+    const auto delta = wrapper::vdup_n(static_cast<float16_t>((1e-7), ExactTagType {});
+#endif /* __aarch64 */
+
+                                       const auto const_1     = wrapper::vdup_n(static_cast<float16_t>(1.f), ExactTagType {});
+                                       const auto const_0     = wrapper::vdup_n(static_cast<float16_t>(0.f), ExactTagType{});
+                                       const auto const_6     = wrapper::vdup_n(static_cast<float16_t>(6.f), ExactTagType{});
+                                       const auto const_3     = wrapper::vdup_n(static_cast<float16_t>(3.f), ExactTagType{});
+                                       const auto const_inv_6 = wrapper::vdup_n(static_cast<float16_t>(0.166666667f), ExactTagType{});
+
+                                       const auto va = wrapper::vdup_n(static_cast<float16_t>(act_info.a()), ExactTagType{});
+                                       const auto vb = wrapper::vdup_n(static_cast<float16_t>(act_info.b()), ExactTagType{});
+                                       const auto a  = static_cast<float16_t>(act_info.a());
+                                       const auto b  = static_cast<float16_t>(act_info.b());
+                                       execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const float16_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr());
+
+        wrapper::traits::neon_bitvector_t<float16_t, wrapper::traits::BitWidth::W128> tmp;
+
+        // Compute S elements per iteration
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            const auto vin = wrapper::vloadq(input_ptr + x);
+            switch(act)
+            {
+                case ActivationLayerInfo::ActivationFunction::ABS:
+                    tmp = wrapper::vabs(vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LINEAR:
+                    tmp = wrapper::vmla(vb, va, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LOGISTIC:
+                    tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin))));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::RELU:
+                    tmp = wrapper::vmax(const_0, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+                    tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
+                    tmp = wrapper::vmin(va, wrapper::vmax(vb, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+                    tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
+                    tmp = wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::ELU:
+                    tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQRT:
+#ifdef __aarch64__
+                    tmp = wrapper::vsqrt(vin);
+#else  /* aarch64 */
+                    {
+                        const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0, ExactTagType{}));
+                        tmp                 = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask))));
+                        tmp                 = mask_float_vector(tmp, wrapper::vnot(bitmask));
+                    }
+#endif /* aarch64 */
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQUARE:
+                    tmp = wrapper::vmul(vin, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::TANH:
+                    tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::IDENTITY:
+                    tmp = vin;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
+                    tmp = wrapper::vmul(vin, wrapper::vmul(const_inv_6, wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3)))));
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            wrapper::vstore(output_ptr + x, tmp);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            const float16_t in = *(reinterpret_cast<const float16_t *>(input_ptr + x));
+            float16_t       tmp;
+            switch(act)
+            {
+                case ActivationLayerInfo::ActivationFunction::ABS:
+                    tmp = std::abs(in);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LINEAR:
+                    tmp = a * in + b;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LOGISTIC:
+                    tmp = static_cast<float16_t>(1) / (static_cast<float16_t>(1) + std::exp(-in));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::RELU:
+                    tmp = std::max<float16_t>(static_cast<float16_t>(0), in);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+                    tmp = std::min<float16_t>(a, std::max(static_cast<float16_t>(0), in));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
+                    tmp = std::min<float16_t>(a, std::max<float16_t>(b, in));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+                    tmp = (in > 0) ? in : a * in;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
+                    tmp = std::log(static_cast<float16_t>(1) + std::exp(in));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::ELU:
+                    tmp = (in >= 0) ? in : a * (std::exp(in) - 1);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQRT:
+                    tmp = std::sqrt(in);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQUARE:
+                    tmp = in * in;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::TANH:
+                    tmp = a * std::tanh(b * in);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::IDENTITY:
+                    tmp = in;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
+                    tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f);
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            *(output_ptr + x) = tmp;
+        }
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
\ No newline at end of file
diff --git a/src/core/cpu/kernels/activation/NEON/fp32.cpp b/src/core/cpu/kernels/activation/NEON/fp32.cpp
new file mode 100644
index 0000000000..0687646be7
--- /dev/null
+++ b/src/core/cpu/kernels/activation/NEON/fp32.cpp
@@ -0,0 +1,210 @@
+/*
+ * Copyright (c) 2020-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace
+{
+#ifndef __aarch64__
+inline float32x4_t mask_float_vector(const float32x4_t &in, const uint32x4_t &mask)
+{
+    auto int_in = vreinterpretq_u32_f32(in);
+    return vreinterpretq_f32_u32(wrapper::vand(int_in, mask));
+}
+#endif /* __arch64__ */
+} // namespace
+
+void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+{
+    /** NEON vector tag type. */
+    using ExactTagType = typename arm_compute::wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>;
+
+    constexpr int                                 window_step_x  = 4;
+    const auto                                    window_start_x = static_cast<int>(window.x().start());
+    const auto                                    window_end_x   = static_cast<int>(window.x().end());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    // In case of non-aarch64, a small delta value is added to the input
+    // to prevent NAN values caused by zeros in inputs to SQRT.
+    // In case of aarh64, we call vsqrt directly, so we don't use delta.
+#ifndef __aarch64__
+    const auto delta = wrapper::vdup_n(static_cast<float>(1e-24), ExactTagType {});
+#endif /* __aarch64 */
+    const auto const_1     = wrapper::vdup_n(static_cast<float>(1.f), ExactTagType {});
+    const auto const_0     = wrapper::vdup_n(static_cast<float>(0.f), ExactTagType{});
+    const auto const_6     = wrapper::vdup_n(static_cast<float>(6.f), ExactTagType{});
+    const auto const_3     = wrapper::vdup_n(static_cast<float>(3.f), ExactTagType{});
+    const auto const_inv_6 = wrapper::vdup_n(static_cast<float>(0.166666667f), ExactTagType{});
+
+    const auto va = wrapper::vdup_n(static_cast<float>(act_info.a()), ExactTagType{});
+    const auto vb = wrapper::vdup_n(static_cast<float>(act_info.b()), ExactTagType{});
+    const auto a  = static_cast<float>(act_info.a());
+    const auto b  = static_cast<float>(act_info.b());
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const float *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<float *>(output.ptr());
+
+        wrapper::traits::neon_bitvector_t<float, wrapper::traits::BitWidth::W128> tmp;
+
+        // Compute S elements per iteration
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            const auto vin = wrapper::vloadq(input_ptr + x);
+            switch(act)
+            {
+                case ActivationLayerInfo::ActivationFunction::ABS:
+                    tmp = wrapper::vabs(vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LINEAR:
+                    tmp = wrapper::vmla(vb, va, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LOGISTIC:
+                    tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin))));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::RELU:
+                    tmp = wrapper::vmax(const_0, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+                    tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
+                    tmp = wrapper::vmin(va, wrapper::vmax(vb, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+                    tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
+                    tmp = wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::ELU:
+                    tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQRT:
+#ifdef __aarch64__
+                    tmp = wrapper::vsqrt(vin);
+#else  /* aarch64 */
+                    {
+                        const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0.f, ExactTagType{}));
+                        tmp                 = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask))));
+                        tmp                 = mask_float_vector(tmp, wrapper::vnot(bitmask));
+                    }
+#endif /* aarch64 */
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQUARE:
+                    tmp = wrapper::vmul(vin, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::TANH:
+                    tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::IDENTITY:
+                    tmp = vin;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
+                    tmp = wrapper::vmul(vin, wrapper::vmul(const_inv_6, wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3)))));
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            wrapper::vstore(output_ptr + x, tmp);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            const float in = *(reinterpret_cast<const float *>(input_ptr + x));
+            float       tmp;
+            switch(act)
+            {
+                case ActivationLayerInfo::ActivationFunction::ABS:
+                    tmp = std::abs(in);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LINEAR:
+                    tmp = a * in + b;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LOGISTIC:
+                    tmp = static_cast<float>(1) / (static_cast<float>(1) + std::exp(-in));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::RELU:
+                    tmp = std::max<float>(static_cast<float>(0), in);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+                    tmp = std::min<float>(a, std::max(static_cast<float>(0), in));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
+                    tmp = std::min<float>(a, std::max<float>(b, in));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+                    tmp = (in > 0) ? in : a * in;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
+                    tmp = std::log(static_cast<float>(1) + std::exp(in));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::ELU:
+                    tmp = (in >= 0) ? in : a * (std::exp(in) - 1);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQRT:
+                    tmp = std::sqrt(in);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQUARE:
+                    tmp = in * in;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::TANH:
+                    tmp = a * std::tanh(b * in);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::IDENTITY:
+                    tmp = in;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
+                    tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f);
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            *(output_ptr + x) = tmp;
+        }
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/activation/NEON/qasymm8.cpp b/src/core/cpu/kernels/activation/NEON/qasymm8.cpp
new file mode 100644
index 0000000000..7506a8294f
--- /dev/null
+++ b/src/core/cpu/kernels/activation/NEON/qasymm8.cpp
@@ -0,0 +1,263 @@
+/*
+ * Copyright (c) 2020-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEAsymm.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void qasymm8_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+{
+    constexpr int                                 window_step_x  = 16;
+    const auto                                    window_start_x = static_cast<int>(window.x().start());
+    const auto                                    window_end_x   = static_cast<int>(window.x().end());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const UniformQuantizationInfo qi_in    = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qi_out   = dst->info()->quantization_info().uniform();
+    const qasymm8x16_t            va       = vdupq_n_u8(quantize_qasymm8(act_info.a(), qi_in));
+    const qasymm8x16_t            vb       = vdupq_n_u8(quantize_qasymm8(act_info.b(), qi_in));
+    const qasymm8_t               a        = quantize_qasymm8(act_info.a(), qi_in);
+    const qasymm8_t               b        = quantize_qasymm8(act_info.b(), qi_in);
+    const qasymm8_t               const_0  = quantize_qasymm8(0.f, qi_in);
+    const qasymm8x16_t            vconst_0 = vdupq_n_u8(const_0);
+    const auto                    vconst_1 = vdupq_n_f32(1.f);
+#ifndef __aarch64__
+    const auto vconst_0_f32 = vdupq_n_f32(0);
+#endif // __aarch64__
+    const float32x4_t va_f32          = vdupq_n_f32(act_info.a());
+    const float32x4_t vb_f32          = vdupq_n_f32(act_info.b());
+    const float       a_f32           = act_info.a();
+    const float       b_f32           = act_info.b();
+    const auto        const_6_f32     = vdupq_n_f32(6.f);
+    const auto        const_0_f32     = vdupq_n_f32(0.f);
+    const auto        const_3_f32     = vdupq_n_f32(3.f);
+    const auto        const_inv_6_f32 = vdupq_n_f32(0.166666667f);
+
+    // Initialise scale/offset for re-quantization
+    float       s  = qi_in.scale / qi_out.scale;
+    float       o  = -qi_in.offset * s + qi_out.offset;
+    float32x4_t vs = vdupq_n_f32(s);
+    float32x4_t vo = vdupq_n_f32(o);
+
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const qasymm8_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<qasymm8_t *>(output.ptr());
+
+        wrapper::traits::neon_bitvector_t<qasymm8_t, wrapper::traits::BitWidth::W128> tmp;
+
+        // Compute S elements per iteration
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            const auto vin = wrapper::vloadq(input_ptr + x);
+            if(act == ActivationLayerInfo::ActivationFunction::RELU)
+            {
+                // Perform activation
+                tmp = vmaxq_u8(vconst_0, vin);
+                // Re-quantize to new output space
+                tmp = vmlaq_qasymm8(tmp, vs, vo);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = vminq_u8(va, vmaxq_u8(vconst_0, vin));
+                // Re-quantize to new output space
+                tmp = vmlaq_qasymm8(tmp, vs, vo);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = vminq_u8(va, vmaxq_u8(vb, vin));
+                // Re-quantize to new output space
+                tmp = vmlaq_qasymm8(tmp, vs, vo);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize(vin, qi_in);
+                // Perform activation
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize(tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::TANH)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize(vin, qi_in);
+                // Perform activation
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize(tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize(vin, qi_in);
+                // Perform activation
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vmul(vin_deq.val[0], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))),
+                        wrapper::vmul(vin_deq.val[1], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))),
+                        wrapper::vmul(vin_deq.val[2], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))),
+                        wrapper::vmul(vin_deq.val[3], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize(tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
+            {
+                const auto vin_deq = vdequantize(vin, qi_in);
+
+#ifdef __aarch64__
+                const uint32x4x4_t pos_mask =
+                {
+                    {
+                        wrapper::vcgtz(vin_deq.val[0]),
+                        wrapper::vcgtz(vin_deq.val[1]),
+                        wrapper::vcgtz(vin_deq.val[2]),
+                        wrapper::vcgtz(vin_deq.val[3]),
+                    }
+                };
+#else  // __aarch64__
+                const uint32x4x4_t pos_mask =
+                {
+                    {
+                        wrapper::vcgt(vin_deq.val[0], vconst_0_f32),
+                        wrapper::vcgt(vin_deq.val[1], vconst_0_f32),
+                        wrapper::vcgt(vin_deq.val[2], vconst_0_f32),
+                        wrapper::vcgt(vin_deq.val[3], vconst_0_f32),
+                    }
+                };
+#endif // __aarch64__
+
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vbsl(pos_mask.val[0], vin_deq.val[0], wrapper::vmul(va_f32, vin_deq.val[0])),
+                        wrapper::vbsl(pos_mask.val[1], vin_deq.val[1], wrapper::vmul(va_f32, vin_deq.val[1])),
+                        wrapper::vbsl(pos_mask.val[2], vin_deq.val[2], wrapper::vmul(va_f32, vin_deq.val[2])),
+                        wrapper::vbsl(pos_mask.val[3], vin_deq.val[3], wrapper::vmul(va_f32, vin_deq.val[3])),
+                    }
+                };
+
+                tmp = vquantize(tmp_dep, qi_out);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            wrapper::vstore(output_ptr + x, tmp);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            qasymm8_t in  = *(reinterpret_cast<const qasymm8_t *>(input_ptr + x));
+            qasymm8_t tmp = 0;
+            if(act == ActivationLayerInfo::ActivationFunction::RELU)
+            {
+                tmp = std::max(const_0, in);
+                tmp = utility::clamp<int32_t, qasymm8_t>(tmp * s + o);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+            {
+                tmp = std::min(a, std::max(const_0, in));
+                tmp = utility::clamp<int32_t, qasymm8_t>(tmp * s + o);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+            {
+                tmp = std::min(a, std::max(b, in));
+                tmp = utility::clamp<int32_t, qasymm8_t>(tmp * s + o);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+            {
+                float tmp_f = dequantize_qasymm8(in, qi_in);
+                tmp_f       = 1.f / (1.f + std::exp(-tmp_f));
+                tmp         = quantize_qasymm8(tmp_f, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::TANH)
+            {
+                float tmp_f = dequantize_qasymm8(in, qi_in);
+                tmp_f       = a_f32 * std::tanh(b_f32 * tmp_f);
+                tmp         = quantize_qasymm8(tmp_f, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
+            {
+                float tmp_f = dequantize_qasymm8(in, qi_in);
+                tmp_f       = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f);
+                tmp         = quantize_qasymm8(tmp_f, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
+            {
+                float tmp_f = dequantize_qasymm8(in, qi_in);
+                tmp_f       = tmp_f > 0 ? tmp_f : tmp_f * a_f32;
+                tmp         = quantize_qasymm8(tmp_f, qi_out);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            *(output_ptr + x) = tmp;
+        }
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/activation/NEON/qasymm8_signed.cpp b/src/core/cpu/kernels/activation/NEON/qasymm8_signed.cpp
new file mode 100644
index 0000000000..8f75abea8e
--- /dev/null
+++ b/src/core/cpu/kernels/activation/NEON/qasymm8_signed.cpp
@@ -0,0 +1,262 @@
+/*
+ * Copyright (c) 2020-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEAsymm.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void qasymm8_signed_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+{
+    constexpr int                                 window_step_x  = 16;
+    const auto                                    window_start_x = static_cast<int>(window.x().start());
+    const auto                                    window_end_x   = static_cast<int>(window.x().end());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const UniformQuantizationInfo qi_in    = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qi_out   = dst->info()->quantization_info().uniform();
+    const qasymm8x16_signed_t     va       = vdupq_n_s8(quantize_qasymm8_signed(act_info.a(), qi_in));
+    const qasymm8x16_signed_t     vb       = vdupq_n_s8(quantize_qasymm8_signed(act_info.b(), qi_in));
+    const qasymm8_signed_t        a        = quantize_qasymm8_signed(act_info.a(), qi_in);
+    const qasymm8_signed_t        b        = quantize_qasymm8_signed(act_info.b(), qi_in);
+    const qasymm8_signed_t        const_0  = quantize_qasymm8_signed(0.f, qi_in);
+    const qasymm8x16_signed_t     vconst_0 = vdupq_n_s8(const_0);
+    const auto                    vconst_1 = vdupq_n_f32(1.f);
+#ifndef __aarch64__
+    const auto vconst_0_f32 = vdupq_n_f32(1.f);
+#endif // __aarch64__
+    const float32x4_t va_f32          = vdupq_n_f32(act_info.a());
+    const float32x4_t vb_f32          = vdupq_n_f32(act_info.b());
+    const float       a_f32           = act_info.a();
+    const float       b_f32           = act_info.b();
+    const auto        const_6_f32     = vdupq_n_f32(6.f);
+    const auto        const_0_f32     = vdupq_n_f32(0.f);
+    const auto        const_3_f32     = vdupq_n_f32(3.f);
+    const auto        const_inv_6_f32 = vdupq_n_f32(0.166666667f);
+
+    // Initialise scale/offset for re-quantization
+    float       s  = qi_in.scale / qi_out.scale;
+    float       o  = -qi_in.offset * s + qi_out.offset;
+    float32x4_t vs = vdupq_n_f32(s);
+    float32x4_t vo = vdupq_n_f32(o);
+
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const qasymm8_signed_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<qasymm8_signed_t *>(output.ptr());
+
+        wrapper::traits::neon_bitvector_t<qasymm8_signed_t, wrapper::traits::BitWidth::W128> tmp;
+
+        // Compute S elements per iteration
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            const auto vin = wrapper::vloadq(input_ptr + x);
+            if(act == ActivationLayerInfo::ActivationFunction::RELU)
+            {
+                // Perform activation
+                tmp = vmaxq_s8(vconst_0, vin);
+                // Re-quantize to new output space
+                tmp = vmlaq_qasymm8_signed(tmp, vs, vo);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = vminq_s8(va, vmaxq_s8(vconst_0, vin));
+                // Re-quantize to new output space
+                tmp = vmlaq_qasymm8_signed(tmp, vs, vo);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = vminq_s8(va, vmaxq_s8(vb, vin));
+                // Re-quantize to new output space
+                tmp = vmlaq_qasymm8_signed(tmp, vs, vo);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize(vin, qi_in);
+                // Perform activation
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize_signed(tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::TANH)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize(vin, qi_in);
+                // Perform activation
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize_signed(tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize(vin, qi_in);
+                // Perform activation
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vmul(vin_deq.val[0], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))),
+                        wrapper::vmul(vin_deq.val[1], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))),
+                        wrapper::vmul(vin_deq.val[2], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))),
+                        wrapper::vmul(vin_deq.val[3], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize_signed(tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
+            {
+                const auto vin_deq = vdequantize(vin, qi_in);
+
+#ifdef __aarch64__
+                const uint32x4x4_t pos_mask =
+                {
+                    {
+                        wrapper::vcgtz(vin_deq.val[0]),
+                        wrapper::vcgtz(vin_deq.val[1]),
+                        wrapper::vcgtz(vin_deq.val[2]),
+                        wrapper::vcgtz(vin_deq.val[3]),
+                    }
+                };
+#else  // __aarch64__
+                const uint32x4x4_t pos_mask =
+                {
+                    {
+                        wrapper::vcgt(vin_deq.val[0], vconst_0_f32),
+                        wrapper::vcgt(vin_deq.val[1], vconst_0_f32),
+                        wrapper::vcgt(vin_deq.val[2], vconst_0_f32),
+                        wrapper::vcgt(vin_deq.val[3], vconst_0_f32),
+                    }
+                };
+#endif // __aarch64__
+
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vbsl(pos_mask.val[0], vin_deq.val[0], wrapper::vmul(va_f32, vin_deq.val[0])),
+                        wrapper::vbsl(pos_mask.val[1], vin_deq.val[1], wrapper::vmul(va_f32, vin_deq.val[1])),
+                        wrapper::vbsl(pos_mask.val[2], vin_deq.val[2], wrapper::vmul(va_f32, vin_deq.val[2])),
+                        wrapper::vbsl(pos_mask.val[3], vin_deq.val[3], wrapper::vmul(va_f32, vin_deq.val[3])),
+                    }
+                };
+
+                tmp = vquantize_signed(tmp_dep, qi_out);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            wrapper::vstore(output_ptr + x, tmp);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            qasymm8_signed_t in  = *(reinterpret_cast<const qasymm8_signed_t *>(input_ptr + x));
+            qasymm8_signed_t tmp = 0;
+            if(act == ActivationLayerInfo::ActivationFunction::RELU)
+            {
+                tmp = std::max(const_0, in);
+                tmp = utility::clamp<int32_t, qasymm8_signed_t>(tmp * s + o);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+            {
+                tmp = std::min(a, std::max(const_0, in));
+                tmp = utility::clamp<int32_t, qasymm8_signed_t>(tmp * s + o);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+            {
+                tmp = std::min(a, std::max(b, in));
+                tmp = utility::clamp<int32_t, qasymm8_signed_t>(tmp * s + o);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+            {
+                float tmp_f = dequantize_qasymm8_signed(in, qi_in);
+                tmp_f       = 1.f / (1.f + std::exp(-tmp_f));
+                tmp         = quantize_qasymm8_signed(tmp_f, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::TANH)
+            {
+                float tmp_f = dequantize_qasymm8_signed(in, qi_in);
+                tmp_f       = a_f32 * std::tanh(b_f32 * tmp_f);
+                tmp         = quantize_qasymm8_signed(tmp_f, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
+            {
+                float tmp_f = dequantize_qasymm8_signed(in, qi_in);
+                tmp_f       = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f);
+                tmp         = quantize_qasymm8_signed(tmp_f, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
+            {
+                float tmp_f = dequantize_qasymm8_signed(in, qi_in);
+                tmp_f       = tmp_f > 0 ? tmp_f : tmp_f * a_f32;
+                tmp         = quantize_qasymm8_signed(tmp_f, qi_out);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            *(output_ptr + x) = tmp;
+        }
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/activation/NEON/qsymm16.cpp b/src/core/cpu/kernels/activation/NEON/qsymm16.cpp
new file mode 100644
index 0000000000..9eee360427
--- /dev/null
+++ b/src/core/cpu/kernels/activation/NEON/qsymm16.cpp
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 2020-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/experimental/Types.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/NESymm.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Validate.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void qsymm16_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+{
+    constexpr int                                 window_step_x  = 8;
+    const auto                                    window_start_x = static_cast<int>(window.x().start());
+    const auto                                    window_end_x   = static_cast<int>(window.x().end());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const UniformQuantizationInfo qi_in    = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qi_out   = dst->info()->quantization_info().uniform();
+    const auto                    vconst_1 = vdupq_n_f32(1.f);
+    const float32x4_t             va_f32   = vdupq_n_f32(act_info.a());
+    const float32x4_t             vb_f32   = vdupq_n_f32(act_info.b());
+    const float                   a_f32    = act_info.a();
+    const float                   b_f32    = act_info.b();
+
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const qsymm16_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<qsymm16_t *>(output.ptr());
+
+        wrapper::traits::neon_bitvector_t<qsymm16_t, wrapper::traits::BitWidth::W128> tmp;
+        ARM_COMPUTE_UNUSED(tmp);
+
+        // Compute S elements per iteration
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            const auto vin = wrapper::vloadq(input_ptr + x);
+            if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize_int16(vin, qi_in.scale);
+                // Perform activation
+                const float32x4x2_t tmp_dep =
+                {
+                    {
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize_int16(tmp_dep, qi_out.scale);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::TANH)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize_int16(vin, qi_in.scale);
+                // Perform activation
+                const float32x4x2_t tmp_dep =
+                {
+                    {
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize_int16(tmp_dep, qi_out.scale);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            wrapper::vstore(output_ptr + x, tmp);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            qsymm16_t in  = *(reinterpret_cast<const qsymm16_t *>(input_ptr + x));
+            qsymm16_t tmp = 0;
+            if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+            {
+                float tmp_f = dequantize_qsymm16(in, qi_in.scale);
+                tmp_f       = 1.f / (1.f + std::exp(-tmp_f));
+                tmp         = quantize_qsymm16(tmp_f, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::TANH)
+            {
+                float tmp_f = dequantize_qsymm16(in, qi_in.scale);
+                tmp_f       = a_f32 * std::tanh(b_f32 * tmp_f);
+                tmp         = quantize_qsymm16(tmp_f, qi_out);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            *(output_ptr + x) = tmp;
+        }
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute