Decouple CpuActivationKernel

1- Data types were already decoupled. This commit arrange the folder struct of the activation kernel.
2- Refactor NEON CpuActivationKernel for floating-point cases.

Resolves COMPMID-4636
Change-Id: Ia4527244c84260dce1dd1d4bd4a9e3cfe2486d85
Signed-off-by: Dana Zlotnik <dana.zlotnik@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6739
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>

16 files changed, 176 insertions, 315 deletions

diff --git a/Android.bp b/Android.bp
index 5727706c07..5654b840fe 100644
--- a/Android.bp
+++ b/Android.bp
@@ -429,16 +429,16 @@ cc_library_static {
         "src/cpu/kernels/CpuTransposeKernel.cpp",
         "src/cpu/kernels/CpuWeightsReshapeKernel.cpp",
         "src/cpu/kernels/CpuWinogradConv2dKernel.cpp",
-        "src/cpu/kernels/activation/neon/fp16.cpp",
-        "src/cpu/kernels/activation/neon/fp32.cpp",
-        "src/cpu/kernels/activation/neon/qasymm8.cpp",
-        "src/cpu/kernels/activation/neon/qasymm8_signed.cpp",
-        "src/cpu/kernels/activation/neon/qsymm16.cpp",
-        "src/cpu/kernels/activation/sve/fp16.cpp",
-        "src/cpu/kernels/activation/sve/fp32.cpp",
-        "src/cpu/kernels/activation/sve/qasymm8.cpp",
-        "src/cpu/kernels/activation/sve/qasymm8_signed.cpp",
-        "src/cpu/kernels/activation/sve/qsymm16.cpp",
+        "src/cpu/kernels/activation/generic/neon/fp16.cpp",
+        "src/cpu/kernels/activation/generic/neon/fp32.cpp",
+        "src/cpu/kernels/activation/generic/neon/qasymm8.cpp",
+        "src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp",
+        "src/cpu/kernels/activation/generic/neon/qsymm16.cpp",
+        "src/cpu/kernels/activation/generic/sve/fp16.cpp",
+        "src/cpu/kernels/activation/generic/sve/fp32.cpp",
+        "src/cpu/kernels/activation/generic/sve2/qasymm8.cpp",
+        "src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp",
+        "src/cpu/kernels/activation/generic/sve2/qsymm16.cpp",
         "src/cpu/kernels/add/generic/neon/fp16.cpp",
         "src/cpu/kernels/add/generic/neon/fp32.cpp",
         "src/cpu/kernels/add/generic/neon/impl.cpp",
diff --git a/filelist.json b/filelist.json
index 428ad7d2cb..d7847480a4 100644
--- a/filelist.json
+++ b/filelist.json
@@ -845,20 +845,20 @@
             "src/cpu/operators/CpuActivation.cpp",
             "src/cpu/kernels/CpuActivationKernel.cpp",
             "src/runtime/NEON/functions/NEActivationLayer.cpp",
-            "src/cpu/kernels/activation/neon/qasymm8.cpp",
-            "src/cpu/kernels/activation/neon/qasymm8_signed.cpp",
-            "src/cpu/kernels/activation/neon/qsymm16.cpp"
+            "src/cpu/kernels/activation/generic/neon/qasymm8.cpp",
+            "src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp",
+            "src/cpu/kernels/activation/generic/neon/qsymm16.cpp"
           ],
           "neon": {
-            "fp16": [ "src/cpu/kernels/activation/neon/fp16.cpp" ],
-            "fp32": [ "src/cpu/kernels/activation/neon/fp32.cpp" ]
+            "fp16": [ "src/cpu/kernels/activation/generic/neon/fp16.cpp" ],
+            "fp32": [ "src/cpu/kernels/activation/generic/neon/fp32.cpp" ]
           },
           "sve": {
-            "fp16": [ "src/cpu/kernels/activation/sve/fp16.cpp" ],
-            "fp32": [ "src/cpu/kernels/activation/sve/fp32.cpp" ],
-            "qasymm8": [ "src/cpu/kernels/activation/sve/qasymm8.cpp" ],
-            "qasymm8_signed": [ "src/cpu/kernels/activation/sve/qasymm8_signed.cpp" ],
-            "qsymm16": [ "src/cpu/kernels/activation/sve/qsymm16.cpp" ]
+            "fp16": [ "src/cpu/kernels/activation/generic/sve/fp16.cpp" ],
+            "fp32": [ "src/cpu/kernels/activation/generic/sve/fp32.cpp" ],
+            "qasymm8": [ "src/cpu/kernels/activation/generic/sve2/qasymm8.cpp" ],
+            "qasymm8_signed": [ "src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp" ],
+            "qsymm16": [ "src/cpu/kernels/activation/generic/sve2/qsymm16.cpp" ]
           }
         }
       },
diff --git a/src/cpu/kernels/CpuActivationKernel.cpp b/src/cpu/kernels/CpuActivationKernel.cpp
index aed73d1fec..3af379d8af 100644
--- a/src/cpu/kernels/CpuActivationKernel.cpp
+++ b/src/cpu/kernels/CpuActivationKernel.cpp
@@ -65,57 +65,57 @@ static const ActivationKernel available_kernels[] =
     {
         "sve_fp16_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::F16 && data.ci.has_sve(); },
-        REGISTER_FP16_SVE(arm_compute::cpu::fp16_sve_activation)
+        REGISTER_FP16_SVE(arm_compute::cpu::sve_fp16_activation)
     },
     {
         "sve_fp32_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::F32 && data.ci.has_sve(); },
-        REGISTER_FP32_SVE(arm_compute::cpu::fp32_sve_activation)
+        REGISTER_FP32_SVE(arm_compute::cpu::sve_fp32_activation)
     },
 #endif /* defined(ARM_COMPUTE_ENABLE_SVE)  */
 #if defined(ARM_COMPUTE_ENABLE_NEON)
     {
         "neon_fp16_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::F16; },
-        REGISTER_FP16_NEON(arm_compute::cpu::fp16_neon_activation)
+        REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_activation)
     },
     {
         "neon_fp32_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::F32; },
-        REGISTER_FP32_NEON(arm_compute::cpu::fp32_neon_activation)
+        REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_activation)
     },
 #endif /* defined(ARM_COMPUTE_ENABLE_NEON)  */
 #if defined(ARM_COMPUTE_ENABLE_SVE2)
     {
         "sve_qu8_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8 && data.ci.has_sve2(); },
-        REGISTER_QASYMM8_SVE(arm_compute::cpu::qasymm8_sve_activation)
+        REGISTER_QASYMM8_SVE2(arm_compute::cpu::sve2_qasymm8_activation)
     },
     {
         "sve_qs8_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED && data.ci.has_sve2(); },
-        REGISTER_QASYMM8_SIGNED_SVE(arm_compute::cpu::qasymm8_signed_sve_activation)
+        REGISTER_QASYMM8_SIGNED_SVE2(arm_compute::cpu::sve2_qasymm8_signed_activation)
     },
     {
         "sve_qs16_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::QSYMM16 && data.ci.has_sve2(); },
-        REGISTER_QSYMM16_SVE(arm_compute::cpu::qsymm16_sve_activation)
+        REGISTER_QSYMM16_SVE2(arm_compute::cpu::sve2_qsymm16_activation)
     },
 #endif /* defined(ARM_COMPUTE_ENABLE_SVE2) */
     {
         "neon_qu8_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8; },
-        REGISTER_QASYMM8_NEON(arm_compute::cpu::qasymm8_neon_activation)
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_activation)
     },
     {
         "neon_qs8_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
-        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::qasymm8_signed_neon_activation)
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::neon_qasymm8_signed_activation)
     },
     {
         "neon_qs16_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::QSYMM16; },
-        REGISTER_QSYMM16_NEON(arm_compute::cpu::qsymm16_neon_activation)
+        REGISTER_QSYMM16_NEON(arm_compute::cpu::neon_qsymm16_activation)
     },
 };
 
@@ -233,18 +233,14 @@ Status CpuActivationKernel::validate(const ITensorInfo *src, const ITensorInfo *
 size_t CpuActivationKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
 {
     ARM_COMPUTE_UNUSED(thread_count);
-    // Tuning results that gave optimized results in performance investigation 
-    if (platform.get_cpu_model() == CPUModel::A73 ) 
+    // Tuning results that gave optimized results in performance investigation
+    if(platform.get_cpu_model() == CPUModel::A73)
     {
         return 10240;
     }
-    else if (platform.get_cpu_model() == CPUModel::A76)
-    {
-        return 9216;
-    }
     else
     {
-        return ICPPKernel::default_mws;
+        return 9216;
     }
 }
 
diff --git a/src/cpu/kernels/activation/generic/neon/fp16.cpp b/src/cpu/kernels/activation/generic/neon/fp16.cpp
new file mode 100644
index 0000000000..e51b5b3423
--- /dev/null
+++ b/src/cpu/kernels/activation/generic/neon/fp16.cpp
@@ -0,0 +1,43 @@
+/*
+ * 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.
+ */
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+
+#include "src/cpu/kernels/activation/generic/neon/impl.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace
+{
+constexpr ActFpImplParams Fp16Params = { static_cast<float16_t>(1e-7), 8 };
+} // namespace
+
+void neon_fp16_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+{
+    fp_neon_activation_impl<float16_t, Fp16Params>(src, dst, act_info, window);
+}
+} // 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/cpu/kernels/activation/generic/neon/fp32.cpp b/src/cpu/kernels/activation/generic/neon/fp32.cpp
new file mode 100644
index 0000000000..2a3b8a0bfd
--- /dev/null
+++ b/src/cpu/kernels/activation/generic/neon/fp32.cpp
@@ -0,0 +1,39 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "src/cpu/kernels/activation/generic/neon/impl.h"
+namespace arm_compute
+{
+namespace cpu
+{
+namespace
+{
+constexpr ActFpImplParams Fp32Params = { static_cast<float>(1e-24), 4 };
+} // namespace
+void neon_fp32_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+{
+    fp_neon_activation_impl<float, Fp32Params>(src, dst, act_info, window);
+}
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/cpu/kernels/activation/neon/fp32.cpp b/src/cpu/kernels/activation/generic/neon/impl.h
index 54301d45ad..2dd239e3a1 100644
--- a/src/cpu/kernels/activation/neon/fp32.cpp
+++ b/src/cpu/kernels/activation/generic/neon/impl.h
@@ -22,72 +22,73 @@
  * 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 <arm_neon.h>
-#include <cmath>
-#include <cstddef>
-
 namespace arm_compute
 {
 namespace cpu
 {
-namespace
+/** Constant parameters needed by the activation implementation.
+ *  These parameters differ for each floating type
+ *
+ * @note This are passed as a struct as C++ does not allow float as a template parameter until C++20
+ **/
+struct ActFpImplParams
 {
+    float delta;  /**< Minimum delta needed to avoid NaN on corner-cases of elementary functions */
+    int   step_x; /**< Window step at the x dimension */
+};
+
 #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));
 }
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+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 //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
 #endif /* __aarch64__ */
-} // namespace
 
-void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+template <typename T, const ActFpImplParams &P>
+void fp_neon_activation_impl(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 {
     /** SIMD 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;
+    using ExactTagType                                           = typename arm_compute::wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
+    constexpr int                                 window_step_x  = P.step_x;
     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);
+    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 {});
+    const auto delta = wrapper::vdup_n(static_cast<T>(P.delta), 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      const_1           = wrapper::vdup_n(static_cast<T>(1.f), ExactTagType {});
+    const auto      const_0           = wrapper::vdup_n(static_cast<T>(0.f), ExactTagType{});
+    const auto      const_6           = wrapper::vdup_n(static_cast<T>(6.f), ExactTagType{});
+    const auto      const_3           = wrapper::vdup_n(static_cast<T>(3.f), ExactTagType{});
+    const auto      const_inv_6       = wrapper::vdup_n(static_cast<T>(0.166666667f), ExactTagType{});
     constexpr float soft_relu_thresh  = 12.f;
-    const auto      vsoft_relu_thresh = wrapper::vdup_n(static_cast<float>(soft_relu_thresh), 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());
+    const auto      vsoft_relu_thresh = wrapper::vdup_n(static_cast<T>(soft_relu_thresh), ExactTagType{});
+    const auto      va                = wrapper::vdup_n(static_cast<T>(act_info.a()), ExactTagType{});
+    const auto      vb                = wrapper::vdup_n(static_cast<T>(act_info.b()), ExactTagType{});
+    const auto      a                 = static_cast<T>(act_info.a());
+    const auto      b                 = static_cast<T>(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;
-
+        const auto input_ptr  = reinterpret_cast<const T *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<T *>(output.ptr());
+        wrapper::traits::neon_bitvector_t<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)
@@ -150,12 +151,11 @@ void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLaye
             }
             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;
+            const T in = *(reinterpret_cast<const T *>(input_ptr + x));
+            T       tmp;
             switch(act)
             {
                 case ActivationLayerInfo::ActivationFunction::ABS:
@@ -165,22 +165,22 @@ void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLaye
                     tmp = a * in + b;
                     break;
                 case ActivationLayerInfo::ActivationFunction::LOGISTIC:
-                    tmp = static_cast<float>(1) / (static_cast<float>(1) + std::exp(-in));
+                    tmp = static_cast<T>(1) / (static_cast<T>(1) + std::exp(-in));
                     break;
                 case ActivationLayerInfo::ActivationFunction::RELU:
-                    tmp = std::max<float>(static_cast<float>(0), in);
+                    tmp = std::max<T>(static_cast<T>(0), in);
                     break;
                 case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
-                    tmp = std::min<float>(a, std::max(static_cast<float>(0), in));
+                    tmp = std::min<T>(a, std::max(static_cast<T>(0), in));
                     break;
                 case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
-                    tmp = std::min<float>(a, std::max<float>(b, in));
+                    tmp = std::min<T>(a, std::max<T>(b, in));
                     break;
                 case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
                     tmp = (in > 0) ? in : a * in;
                     break;
                 case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
-                    tmp = (in > soft_relu_thresh) ? in : std::log(static_cast<float>(1) + std::exp(in));
+                    tmp = (in > soft_relu_thresh) ? in : std::log(static_cast<T>(1) + std::exp(in));
                     break;
                 case ActivationLayerInfo::ActivationFunction::ELU:
                     tmp = (in >= 0) ? in : a * (std::exp(in) - 1);
diff --git a/src/cpu/kernels/activation/neon/qasymm8.cpp b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp
index a1217435b6..62e329e691 100644
--- a/src/cpu/kernels/activation/neon/qasymm8.cpp
+++ b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp
@@ -36,7 +36,7 @@ namespace arm_compute
 {
 namespace cpu
 {
-void qasymm8_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+void neon_qasymm8_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());
diff --git a/src/cpu/kernels/activation/neon/qasymm8_signed.cpp b/src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp
index 8b40bf8e72..4dca1ba794 100644
--- a/src/cpu/kernels/activation/neon/qasymm8_signed.cpp
+++ b/src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp
@@ -35,7 +35,7 @@ namespace arm_compute
 {
 namespace cpu
 {
-void qasymm8_signed_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+void neon_qasymm8_signed_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());
diff --git a/src/cpu/kernels/activation/neon/qsymm16.cpp b/src/cpu/kernels/activation/generic/neon/qsymm16.cpp
index 54b41820f2..865b9f114e 100644
--- a/src/cpu/kernels/activation/neon/qsymm16.cpp
+++ b/src/cpu/kernels/activation/generic/neon/qsymm16.cpp
@@ -37,7 +37,7 @@ namespace arm_compute
 {
 namespace cpu
 {
-void qsymm16_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+void neon_qsymm16_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());
diff --git a/src/cpu/kernels/activation/sve/fp16.cpp b/src/cpu/kernels/activation/generic/sve/fp16.cpp
index 5e76e82c52..47d9fabb55 100644
--- a/src/cpu/kernels/activation/sve/fp16.cpp
+++ b/src/cpu/kernels/activation/generic/sve/fp16.cpp
@@ -36,7 +36,7 @@ namespace arm_compute
 {
 namespace cpu
 {
-void fp16_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+void sve_fp16_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 {
     const auto                                    window_start_x = static_cast<int>(window.x().start());
     const auto                                    window_end_x   = static_cast<int>(window.x().end());
diff --git a/src/cpu/kernels/activation/sve/fp32.cpp b/src/cpu/kernels/activation/generic/sve/fp32.cpp
index cb9f82eb39..1685b0f669 100644
--- a/src/cpu/kernels/activation/sve/fp32.cpp
+++ b/src/cpu/kernels/activation/generic/sve/fp32.cpp
@@ -36,7 +36,7 @@ namespace arm_compute
 {
 namespace cpu
 {
-void fp32_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+void sve_fp32_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 {
     const auto                                    window_start_x = static_cast<int>(window.x().start());
     const auto                                    window_end_x   = static_cast<int>(window.x().end());
diff --git a/src/cpu/kernels/activation/sve/qasymm8.cpp b/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp
index 69fffd96c5..3b99c0f120 100644
--- a/src/cpu/kernels/activation/sve/qasymm8.cpp
+++ b/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp
@@ -36,7 +36,7 @@ namespace arm_compute
 {
 namespace cpu
 {
-void qasymm8_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+void sve2_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 {
     const auto                                    window_start_x = static_cast<int>(window.x().start());
     const auto                                    window_end_x   = static_cast<int>(window.x().end());
diff --git a/src/cpu/kernels/activation/sve/qasymm8_signed.cpp b/src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp
index 53ee515ff9..24415145d3 100644
--- a/src/cpu/kernels/activation/sve/qasymm8_signed.cpp
+++ b/src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp
@@ -37,7 +37,7 @@ namespace arm_compute
 {
 namespace cpu
 {
-void qasymm8_signed_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+void sve2_qasymm8_signed_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 {
     const auto                                    window_start_x = static_cast<int>(window.x().start());
     const auto                                    window_end_x   = static_cast<int>(window.x().end());
diff --git a/src/cpu/kernels/activation/sve/qsymm16.cpp b/src/cpu/kernels/activation/generic/sve2/qsymm16.cpp
index ac549770a2..0eecfa618f 100644
--- a/src/cpu/kernels/activation/sve/qsymm16.cpp
+++ b/src/cpu/kernels/activation/generic/sve2/qsymm16.cpp
@@ -38,7 +38,7 @@ namespace arm_compute
 {
 namespace cpu
 {
-void qsymm16_sve_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+void sve2_qsymm16_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 {
     const auto                                    window_start_x = static_cast<int>(window.x().start());
     const auto                                    window_end_x   = static_cast<int>(window.x().end());
diff --git a/src/cpu/kernels/activation/list.h b/src/cpu/kernels/activation/list.h
index 409d025db0..bf9aa0f373 100644
--- a/src/cpu/kernels/activation/list.h
+++ b/src/cpu/kernels/activation/list.h
@@ -31,16 +31,16 @@ namespace cpu
 #define DECLARE_ACTIVATION_KERNEL(func_name) \
     void func_name(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 
-DECLARE_ACTIVATION_KERNEL(qasymm8_neon_activation);
-DECLARE_ACTIVATION_KERNEL(qasymm8_sve_activation);
-DECLARE_ACTIVATION_KERNEL(qasymm8_signed_neon_activation);
-DECLARE_ACTIVATION_KERNEL(qasymm8_signed_sve_activation);
-DECLARE_ACTIVATION_KERNEL(qsymm16_neon_activation);
-DECLARE_ACTIVATION_KERNEL(qsymm16_sve_activation);
-DECLARE_ACTIVATION_KERNEL(fp16_neon_activation);
-DECLARE_ACTIVATION_KERNEL(fp16_sve_activation);
-DECLARE_ACTIVATION_KERNEL(fp32_neon_activation);
-DECLARE_ACTIVATION_KERNEL(fp32_sve_activation);
+DECLARE_ACTIVATION_KERNEL(neon_qasymm8_activation);
+DECLARE_ACTIVATION_KERNEL(sve2_qasymm8_activation);
+DECLARE_ACTIVATION_KERNEL(neon_qasymm8_signed_activation);
+DECLARE_ACTIVATION_KERNEL(sve2_qasymm8_signed_activation);
+DECLARE_ACTIVATION_KERNEL(neon_qsymm16_activation);
+DECLARE_ACTIVATION_KERNEL(sve2_qsymm16_activation);
+DECLARE_ACTIVATION_KERNEL(sve_fp16_activation);
+DECLARE_ACTIVATION_KERNEL(sve_fp32_activation);
+DECLARE_ACTIVATION_KERNEL(neon_fp16_activation);
+DECLARE_ACTIVATION_KERNEL(neon_fp32_activation);
 
 #undef DECLARE_ACTIVATION_KERNEL
 } // namespace cpu
diff --git a/src/cpu/kernels/activation/neon/fp16.cpp b/src/cpu/kernels/activation/neon/fp16.cpp
deleted file mode 100644
index 6f2d5d8533..0000000000
--- a/src/cpu/kernels/activation/neon/fp16.cpp
+++ /dev/null
@@ -1,217 +0,0 @@
-/*
- * 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 <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 /* __aarch64__ */
-} // namespace
-
-void fp16_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
-{
-    /** SIMD 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{});
-
-    constexpr float soft_relu_thresh  = 12.f;
-    const auto      vsoft_relu_thresh = wrapper::vdup_n(static_cast<float16_t>(soft_relu_thresh), 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::vbsl(wrapper::vcgt(vin, vsoft_relu_thresh), vin, 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 = (in > soft_relu_thresh) ? in : 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) */