Changes to ElementwiseOp to enable fp16 in armv8a multi_isa builds

* Code guarded with __ARM_FEATURE_FP16_VECTOR_ARITHMETIC needs
  to be moved to an fp16.cpp file to allow compilation with
  -march=armv8.2-a+fp16

* Partially resolves MLCE-1102

Change-Id: I5ecfc8f6c0d84f92d80bec2cde6e7338794b9788
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10240
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 250 insertions, 293 deletions

diff --git a/Android.bp b/Android.bp
index b5a0ef9f9a..274d6e7f4d 100644
--- a/Android.bp
+++ b/Android.bp
@@ -501,7 +501,6 @@ cc_library_static {
         "src/cpu/kernels/elementwise_binary/generic/neon/qasymm8_signed.cpp",
         "src/cpu/kernels/elementwise_unary/generic/neon/fp16.cpp",
         "src/cpu/kernels/elementwise_unary/generic/neon/fp32.cpp",
-        "src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp",
         "src/cpu/kernels/elementwise_unary/generic/neon/integer.cpp",
         "src/cpu/kernels/elementwise_unary/generic/neon/q8.cpp",
         "src/cpu/kernels/elementwise_unary/generic/neon/qasymm8.cpp",
diff --git a/filelist.json b/filelist.json
index a3f49d2bae..9a87e857ef 100644
--- a/filelist.json
+++ b/filelist.json
@@ -1445,7 +1445,6 @@
             "src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp"
           ],
           "neon": {
-            "common":["src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp"],
             "integer": ["src/cpu/kernels/elementwise_unary/generic/neon/integer.cpp"],
             "fp32": ["src/cpu/kernels/elementwise_unary/generic/neon/fp32.cpp"],
             "fp16": ["src/cpu/kernels/elementwise_unary/generic/neon/fp16.cpp"],
diff --git a/src/BUILD.bazel b/src/BUILD.bazel
index 0eb27c22ef..27251e67f2 100644
--- a/src/BUILD.bazel
+++ b/src/BUILD.bazel
@@ -753,7 +753,6 @@ filegroup(
 	"cpu/kernels/elementwise_binary/generic/neon/qasymm8_signed.cpp",
 	"cpu/kernels/elementwise_unary/generic/neon/fp16.cpp",
 	"cpu/kernels/elementwise_unary/generic/neon/fp32.cpp",
-	"cpu/kernels/elementwise_unary/generic/neon/impl.cpp",
 	"cpu/kernels/elementwise_unary/generic/neon/integer.cpp",
 	"cpu/kernels/elementwise_unary/generic/neon/q8.cpp",
 	"cpu/kernels/elementwise_unary/generic/neon/qasymm8.cpp",
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 8fffa965b9..cf91a2c0a1 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -745,7 +745,6 @@ target_sources(
 	cpu/kernels/elementwise_binary/generic/neon/qasymm8_signed.cpp
 	cpu/kernels/elementwise_unary/generic/neon/fp16.cpp
 	cpu/kernels/elementwise_unary/generic/neon/fp32.cpp
-	cpu/kernels/elementwise_unary/generic/neon/impl.cpp
 	cpu/kernels/elementwise_unary/generic/neon/integer.cpp
 	cpu/kernels/elementwise_unary/generic/neon/q8.cpp
 	cpu/kernels/elementwise_unary/generic/neon/qasymm8.cpp
diff --git a/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp b/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp
deleted file mode 100644
index 1b449340ba..0000000000
--- a/src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp
+++ /dev/null
@@ -1,286 +0,0 @@
-/*
- * Copyright (c) 2018-2023 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#include "src/cpu/kernels/elementwise_unary/generic/neon/impl.h"
-#include "arm_compute/core/Helpers.h"
-#include "src/core/NEON/NEAsymm.h"
-
-namespace arm_compute
-{
-namespace cpu
-{
-template <typename ScalarType>
-inline ScalarType elementwise_op_scalar_imp(ElementWiseUnary op, const ScalarType &a)
-{
-    switch(op)
-    {
-        case ElementWiseUnary::RSQRT:
-            return 1 / sqrt(a);
-        case ElementWiseUnary::EXP:
-            return std::exp(a);
-        case ElementWiseUnary::NEG:
-            return -a;
-        case ElementWiseUnary::LOG:
-            return std::log(a);
-        case ElementWiseUnary::ABS:
-            return std::abs(a);
-        case ElementWiseUnary::ROUND:
-            return support::cpp11::nearbyint(a);
-        case ElementWiseUnary::SIN:
-            return std::sin(a);
-        default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
-    }
-}
-
-template <typename ScalarType, typename VectorType>
-inline VectorType elementwise_op_imp(ElementWiseUnary op, const VectorType &a)
-{
-    switch(op)
-    {
-        case ElementWiseUnary::RSQRT:
-            return wrapper::vinvsqrt(a);
-        case ElementWiseUnary::EXP:
-            return wrapper::vexpq(a);
-        case ElementWiseUnary::NEG:
-            return wrapper::vneg(a);
-        case ElementWiseUnary::LOG:
-            return wrapper::vlog(a);
-        case ElementWiseUnary::ABS:
-            return wrapper::vabs(a);
-        case ElementWiseUnary::ROUND:
-            return wrapper::vround(a);
-        case ElementWiseUnary::SIN:
-            return wrapper::vsin(a);
-        default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
-    }
-}
-
-template <typename ScalarType>
-void elementwise_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
-{
-    const int  window_step_x  = 16 / sizeof(ScalarType);
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
-
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    Iterator input(in, win);
-    Iterator output(out, win);
-
-    execute_window_loop(win, [&](const Coordinates &)
-    {
-        auto       output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
-        const auto input_ptr  = reinterpret_cast<const ScalarType *>(input.ptr());
-
-        int x = window_start_x;
-        for(; x <= window_end_x - window_step_x; x += window_step_x)
-        {
-            wrapper::vstore(output_ptr + x, elementwise_op_imp<ScalarType>(op, wrapper::vloadq(input_ptr + x)));
-        }
-        for(; x < window_end_x; ++x)
-        {
-            *(output_ptr + x) = elementwise_op_scalar_imp(op, *(input_ptr + x));
-        }
-    },
-    input, output);
-}
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
-template void elementwise_op<__fp16>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
-#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
-template void elementwise_op<float>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
-template void elementwise_op<int32_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
-
-template <>
-void elementwise_op<int8_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
-{
-    const int                     window_step_x     = 16;
-    const auto                    window_start_x    = static_cast<int>(window.x().start());
-    const auto                    window_end_x      = static_cast<int>(window.x().end());
-    const UniformQuantizationInfo qi_in             = in->info()->quantization_info().uniform();
-    const UniformQuantizationInfo qi_out            = out->info()->quantization_info().uniform();
-    const auto                    min_clamped_value = vdupq_n_f32((-128 - qi_out.offset) * qi_out.scale);
-    const auto                    max_clamped_value = vdupq_n_f32((127 - qi_out.offset) * qi_out.scale);
-    Window                        win               = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    Iterator input(in, win);
-    Iterator output(out, win);
-
-    execute_window_loop(win, [&](const Coordinates &)
-    {
-        int8x16_t  vout;
-        auto       output_ptr    = reinterpret_cast<int8_t *>(output.ptr());
-        const auto input_ptr     = reinterpret_cast<const int8_t *>(input.ptr());
-        const auto vconst_0_f32  = vdupq_n_f32(0);
-        auto       clamped_value = (op == ElementWiseUnary::LOG) ? min_clamped_value : max_clamped_value;
-
-        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);
-
-            // De-quantize
-            const auto vin_deq = vdequantize(vin, qi_in);
-
-            // Perform activation
-            float32x4x4_t vtmp_deq =
-            {
-                {
-                    elementwise_op_imp<float>(op, vin_deq.val[0]),
-                    elementwise_op_imp<float>(op, vin_deq.val[1]),
-                    elementwise_op_imp<float>(op, vin_deq.val[2]),
-                    elementwise_op_imp<float>(op, vin_deq.val[3]),
-                }
-            };
-
-            if((op == ElementWiseUnary::LOG) || (op == ElementWiseUnary::RSQRT))
-            {
-                vtmp_deq.val[0] = vbslq_f32(vcleq_f32(vin_deq.val[0], vconst_0_f32), clamped_value, vtmp_deq.val[0]);
-                vtmp_deq.val[1] = vbslq_f32(vcleq_f32(vin_deq.val[1], vconst_0_f32), clamped_value, vtmp_deq.val[1]);
-                vtmp_deq.val[2] = vbslq_f32(vcleq_f32(vin_deq.val[2], vconst_0_f32), clamped_value, vtmp_deq.val[2]);
-                vtmp_deq.val[3] = vbslq_f32(vcleq_f32(vin_deq.val[3], vconst_0_f32), clamped_value, vtmp_deq.val[3]);
-            }
-
-            // Re-quantize to new output space
-            vout = vquantize_signed(vtmp_deq, qi_out);
-            wrapper::vstore(output_ptr + x, vout);
-        }
-        for(; x < window_end_x; ++x)
-        {
-            qasymm8_signed_t in    = *(reinterpret_cast<const qasymm8_signed_t *>(input_ptr + x));
-            qasymm8_signed_t tmp   = 0;
-            float            tmp_f = dequantize_qasymm8_signed(in, qi_in);
-            if(tmp_f <= 0.0)
-            {
-                if(op == ElementWiseUnary::LOG)
-                {
-                    tmp_f = (-128 - qi_out.offset) * qi_out.scale;
-                }
-                else if(op == ElementWiseUnary::RSQRT)
-                {
-                    tmp_f = (127 - qi_out.offset) * qi_out.scale;
-                }
-                else
-                {
-                    tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f);
-                }
-            }
-            else
-            {
-                tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f);
-            }
-            tmp = quantize_qasymm8_signed(tmp_f, qi_out, RoundingPolicy::TO_ZERO); // Set rounding policy TO_ZERO to be compatible with vquantize_signed() used above that follow same policy for armv7a.
-            // For aarch64 LUT is used and rounding to nearest is used
-            *(output_ptr + x) = tmp;
-        }
-    },
-    input, output);
-}
-template <>
-void elementwise_op<uint8_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
-{
-    const int                     window_step_x     = 16;
-    const auto                    window_start_x    = static_cast<int>(window.x().start());
-    const auto                    window_end_x      = static_cast<int>(window.x().end());
-    const UniformQuantizationInfo qi_in             = in->info()->quantization_info().uniform();
-    const UniformQuantizationInfo qi_out            = out->info()->quantization_info().uniform();
-    const auto                    vconst_0_f32      = vdupq_n_f32(0);
-    const auto                    min_clamped_value = vdupq_n_f32((0 - qi_out.offset) * qi_out.scale);
-    const auto                    max_clamped_value = vdupq_n_f32((255 - qi_out.offset) * qi_out.scale);
-    Window                        win               = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    Iterator input(in, win);
-    Iterator output(out, win);
-
-    execute_window_loop(win, [&](const Coordinates &)
-    {
-        uint8x16_t vout;
-        auto       clamped_value = (op == ElementWiseUnary::LOG) ? min_clamped_value : max_clamped_value;
-        auto       output_ptr    = reinterpret_cast<uint8_t *>(output.ptr());
-        const auto input_ptr     = reinterpret_cast<const uint8_t *>(input.ptr());
-        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);
-
-            // De-quantize
-            const auto vin_deq = vdequantize(vin, qi_in);
-
-            // Perform activation
-            float32x4x4_t vtmp_deq =
-            {
-                {
-                    elementwise_op_imp<float>(op, vin_deq.val[0]),
-                    elementwise_op_imp<float>(op, vin_deq.val[1]),
-                    elementwise_op_imp<float>(op, vin_deq.val[2]),
-                    elementwise_op_imp<float>(op, vin_deq.val[3]),
-                }
-            };
-            if((op == ElementWiseUnary::LOG) || (op == ElementWiseUnary::RSQRT))
-            {
-                vtmp_deq.val[0] = vbslq_f32(vcleq_f32(vin_deq.val[0], vconst_0_f32), clamped_value, vtmp_deq.val[0]);
-                vtmp_deq.val[1] = vbslq_f32(vcleq_f32(vin_deq.val[1], vconst_0_f32), clamped_value, vtmp_deq.val[1]);
-                vtmp_deq.val[2] = vbslq_f32(vcleq_f32(vin_deq.val[2], vconst_0_f32), clamped_value, vtmp_deq.val[2]);
-                vtmp_deq.val[3] = vbslq_f32(vcleq_f32(vin_deq.val[3], vconst_0_f32), clamped_value, vtmp_deq.val[3]);
-            }
-
-            // Re-quantize to new output space
-            vout = vquantize(vtmp_deq, qi_out);
-            wrapper::vstore(output_ptr + x, vout);
-        }
-        for(; x < window_end_x; ++x)
-        {
-            qasymm8_t in    = *(reinterpret_cast<const qasymm8_t *>(input_ptr + x));
-            qasymm8_t tmp   = 0;
-            float     tmp_f = dequantize_qasymm8(in, qi_in);
-            if(tmp_f <= 0.0)
-            {
-                if(op == ElementWiseUnary::LOG)
-                {
-                    tmp_f = (0 - qi_out.offset) * qi_out.scale;
-                }
-                else if(op == ElementWiseUnary::RSQRT)
-                {
-                    tmp_f = (255 - qi_out.offset) * qi_out.scale;
-                }
-                else
-                {
-                    tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f);
-                }
-            }
-            else
-            {
-                tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f);
-            }
-            tmp               = quantize_qasymm8(tmp_f, qi_out, RoundingPolicy::TO_ZERO);
-            *(output_ptr + x) = tmp;
-        }
-    },
-    input, output);
-}
-} // namespace cpu
-} // namespace arm_compute
diff --git a/src/cpu/kernels/elementwise_unary/generic/neon/impl.h b/src/cpu/kernels/elementwise_unary/generic/neon/impl.h
index 66b8b5fe45..dbc1dde4fa 100644
--- a/src/cpu/kernels/elementwise_unary/generic/neon/impl.h
+++ b/src/cpu/kernels/elementwise_unary/generic/neon/impl.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2022 Arm Limited.
+ * Copyright (c) 2018-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,6 +26,7 @@
 
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Types.h"
+#include "src/core/NEON/NEAsymm.h"
 #include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
 
 namespace arm_compute
@@ -33,9 +34,255 @@ namespace arm_compute
 namespace cpu
 {
 template <typename ScalarType>
-void elementwise_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
+inline ScalarType elementwise_op_scalar_imp(ElementWiseUnary op, const ScalarType &a)
+{
+    switch(op)
+    {
+        case ElementWiseUnary::RSQRT:
+            return 1 / sqrt(a);
+        case ElementWiseUnary::EXP:
+            return std::exp(a);
+        case ElementWiseUnary::NEG:
+            return -a;
+        case ElementWiseUnary::LOG:
+            return std::log(a);
+        case ElementWiseUnary::ABS:
+            return std::abs(a);
+        case ElementWiseUnary::ROUND:
+            return support::cpp11::nearbyint(a);
+        case ElementWiseUnary::SIN:
+            return std::sin(a);
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+}
+
+template <typename ScalarType, typename VectorType>
+inline VectorType elementwise_op_imp(ElementWiseUnary op, const VectorType &a)
+{
+    switch(op)
+    {
+        case ElementWiseUnary::RSQRT:
+            return wrapper::vinvsqrt(a);
+        case ElementWiseUnary::EXP:
+            return wrapper::vexpq(a);
+        case ElementWiseUnary::NEG:
+            return wrapper::vneg(a);
+        case ElementWiseUnary::LOG:
+            return wrapper::vlog(a);
+        case ElementWiseUnary::ABS:
+            return wrapper::vabs(a);
+        case ElementWiseUnary::ROUND:
+            return wrapper::vround(a);
+        case ElementWiseUnary::SIN:
+            return wrapper::vsin(a);
+        default:
+            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+    }
+}
+
+template <typename ScalarType>
+inline void elementwise_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
+{
+    const int  window_step_x  = 16 / sizeof(ScalarType);
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(in, win);
+    Iterator output(out, win);
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        auto       output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
+        const auto input_ptr  = reinterpret_cast<const ScalarType *>(input.ptr());
+
+        int x = window_start_x;
+        for(; x <= window_end_x - window_step_x; x += window_step_x)
+        {
+            wrapper::vstore(output_ptr + x, elementwise_op_imp<ScalarType>(op, wrapper::vloadq(input_ptr + x)));
+        }
+        for(; x < window_end_x; ++x)
+        {
+            *(output_ptr + x) = elementwise_op_scalar_imp(op, *(input_ptr + x));
+        }
+    },
+    input, output);
+}
+
+template <>
+inline void elementwise_op<int8_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
+{
+    const int                     window_step_x     = 16;
+    const auto                    window_start_x    = static_cast<int>(window.x().start());
+    const auto                    window_end_x      = static_cast<int>(window.x().end());
+    const UniformQuantizationInfo qi_in             = in->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qi_out            = out->info()->quantization_info().uniform();
+    const auto                    min_clamped_value = vdupq_n_f32((-128 - qi_out.offset) * qi_out.scale);
+    const auto                    max_clamped_value = vdupq_n_f32((127 - qi_out.offset) * qi_out.scale);
+    Window                        win               = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(in, win);
+    Iterator output(out, win);
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        int8x16_t  vout;
+        auto       output_ptr    = reinterpret_cast<int8_t *>(output.ptr());
+        const auto input_ptr     = reinterpret_cast<const int8_t *>(input.ptr());
+        const auto vconst_0_f32  = vdupq_n_f32(0);
+        auto       clamped_value = (op == ElementWiseUnary::LOG) ? min_clamped_value : max_clamped_value;
+
+        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);
+
+            // De-quantize
+            const auto vin_deq = vdequantize(vin, qi_in);
+
+            // Perform activation
+            float32x4x4_t vtmp_deq =
+            {
+                {
+                    elementwise_op_imp<float>(op, vin_deq.val[0]),
+                    elementwise_op_imp<float>(op, vin_deq.val[1]),
+                    elementwise_op_imp<float>(op, vin_deq.val[2]),
+                    elementwise_op_imp<float>(op, vin_deq.val[3]),
+                }
+            };
+
+            if((op == ElementWiseUnary::LOG) || (op == ElementWiseUnary::RSQRT))
+            {
+                vtmp_deq.val[0] = vbslq_f32(vcleq_f32(vin_deq.val[0], vconst_0_f32), clamped_value, vtmp_deq.val[0]);
+                vtmp_deq.val[1] = vbslq_f32(vcleq_f32(vin_deq.val[1], vconst_0_f32), clamped_value, vtmp_deq.val[1]);
+                vtmp_deq.val[2] = vbslq_f32(vcleq_f32(vin_deq.val[2], vconst_0_f32), clamped_value, vtmp_deq.val[2]);
+                vtmp_deq.val[3] = vbslq_f32(vcleq_f32(vin_deq.val[3], vconst_0_f32), clamped_value, vtmp_deq.val[3]);
+            }
+
+            // Re-quantize to new output space
+            vout = vquantize_signed(vtmp_deq, qi_out);
+            wrapper::vstore(output_ptr + x, vout);
+        }
+        for(; x < window_end_x; ++x)
+        {
+            qasymm8_signed_t in    = *(reinterpret_cast<const qasymm8_signed_t *>(input_ptr + x));
+            qasymm8_signed_t tmp   = 0;
+            float            tmp_f = dequantize_qasymm8_signed(in, qi_in);
+            if(tmp_f <= 0.0)
+            {
+                if(op == ElementWiseUnary::LOG)
+                {
+                    tmp_f = (-128 - qi_out.offset) * qi_out.scale;
+                }
+                else if(op == ElementWiseUnary::RSQRT)
+                {
+                    tmp_f = (127 - qi_out.offset) * qi_out.scale;
+                }
+                else
+                {
+                    tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f);
+                }
+            }
+            else
+            {
+                tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f);
+            }
+            tmp = quantize_qasymm8_signed(tmp_f, qi_out, RoundingPolicy::TO_ZERO); // Set rounding policy TO_ZERO to be compatible with vquantize_signed() used above that follow same policy for armv7a.
+            // For aarch64 LUT is used and rounding to nearest is used
+            *(output_ptr + x) = tmp;
+        }
+    },
+    input, output);
+}
+template <>
+inline void elementwise_op<uint8_t>(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
+{
+    const int                     window_step_x     = 16;
+    const auto                    window_start_x    = static_cast<int>(window.x().start());
+    const auto                    window_end_x      = static_cast<int>(window.x().end());
+    const UniformQuantizationInfo qi_in             = in->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qi_out            = out->info()->quantization_info().uniform();
+    const auto                    vconst_0_f32      = vdupq_n_f32(0);
+    const auto                    min_clamped_value = vdupq_n_f32((0 - qi_out.offset) * qi_out.scale);
+    const auto                    max_clamped_value = vdupq_n_f32((255 - qi_out.offset) * qi_out.scale);
+    Window                        win               = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(in, win);
+    Iterator output(out, win);
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        uint8x16_t vout;
+        auto       clamped_value = (op == ElementWiseUnary::LOG) ? min_clamped_value : max_clamped_value;
+        auto       output_ptr    = reinterpret_cast<uint8_t *>(output.ptr());
+        const auto input_ptr     = reinterpret_cast<const uint8_t *>(input.ptr());
+        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);
+
+            // De-quantize
+            const auto vin_deq = vdequantize(vin, qi_in);
+
+            // Perform activation
+            float32x4x4_t vtmp_deq =
+            {
+                {
+                    elementwise_op_imp<float>(op, vin_deq.val[0]),
+                    elementwise_op_imp<float>(op, vin_deq.val[1]),
+                    elementwise_op_imp<float>(op, vin_deq.val[2]),
+                    elementwise_op_imp<float>(op, vin_deq.val[3]),
+                }
+            };
+            if((op == ElementWiseUnary::LOG) || (op == ElementWiseUnary::RSQRT))
+            {
+                vtmp_deq.val[0] = vbslq_f32(vcleq_f32(vin_deq.val[0], vconst_0_f32), clamped_value, vtmp_deq.val[0]);
+                vtmp_deq.val[1] = vbslq_f32(vcleq_f32(vin_deq.val[1], vconst_0_f32), clamped_value, vtmp_deq.val[1]);
+                vtmp_deq.val[2] = vbslq_f32(vcleq_f32(vin_deq.val[2], vconst_0_f32), clamped_value, vtmp_deq.val[2]);
+                vtmp_deq.val[3] = vbslq_f32(vcleq_f32(vin_deq.val[3], vconst_0_f32), clamped_value, vtmp_deq.val[3]);
+            }
+
+            // Re-quantize to new output space
+            vout = vquantize(vtmp_deq, qi_out);
+            wrapper::vstore(output_ptr + x, vout);
+        }
+        for(; x < window_end_x; ++x)
+        {
+            qasymm8_t in    = *(reinterpret_cast<const qasymm8_t *>(input_ptr + x));
+            qasymm8_t tmp   = 0;
+            float     tmp_f = dequantize_qasymm8(in, qi_in);
+            if(tmp_f <= 0.0)
+            {
+                if(op == ElementWiseUnary::LOG)
+                {
+                    tmp_f = (0 - qi_out.offset) * qi_out.scale;
+                }
+                else if(op == ElementWiseUnary::RSQRT)
+                {
+                    tmp_f = (255 - qi_out.offset) * qi_out.scale;
+                }
+                else
+                {
+                    tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f);
+                }
+            }
+            else
+            {
+                tmp_f = elementwise_op_scalar_imp<float>(op, tmp_f);
+            }
+            tmp               = quantize_qasymm8(tmp_f, qi_out, RoundingPolicy::TO_ZERO);
+            *(output_ptr + x) = tmp;
+        }
+    },
+    input, output);
+}
 
 } // namespace cpu
 } // namespace arm_compute
 
-#endif // SRC_CORE_NEON_KERNELS_ELEMENTWISE_UNARY_LIST_H
\ No newline at end of file
+#endif // SRC_CORE_NEON_KERNELS_ELEMENTWISE_UNARY_LIST_H