Make Sub kernel and operator stateless

- Rename NEArithmeticSubstractionKernel to CpuSubKernel and move files appropriately

- Add CpuSub under src/runtime/cpu/operators

Partially resolves: COMPMID-4007

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I4754ca9101d82dccacca744be6d069764a9c6b55
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4868
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

7 files changed, 1354 insertions, 0 deletions

diff --git a/src/core/cpu/kernels/CpuSubKernel.cpp b/src/core/cpu/kernels/CpuSubKernel.cpp
new file mode 100644
index 0000000000..a03dcf2353
--- /dev/null
+++ b/src/core/cpu/kernels/CpuSubKernel.cpp
@@ -0,0 +1,251 @@
+/*
+ * 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/core/cpu/kernels/CpuSubKernel.h"
+
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Validate.h"
+#include "src/core/CPP/Validate.h"
+#include "src/core/common/Registrars.h"
+#include "src/core/cpu/kernels/sub/neon/list.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+namespace
+{
+struct SubSelectorData
+{
+    DataType dt1;
+    DataType dt2;
+    DataType dt3;
+};
+
+using SubSelectorPtr = std::add_pointer<bool(const SubSelectorData &data)>::type;
+using SubKernelPtr   = std::add_pointer<void(const ITensor *, const ITensor *, ITensor *, const ConvertPolicy &, const Window &)>::type;
+
+struct SubKernel
+{
+    const char          *name;
+    const SubSelectorPtr is_selected;
+    SubKernelPtr         ukernel;
+};
+
+static const SubKernel available_kernels[] =
+{
+    {
+        "sub_same_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F32)); },
+        REGISTER_FP32_NEON(arm_compute::cpu::sub_same_neon<float>)
+    },
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    {
+        "sub_same_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::F16)); },
+        REGISTER_FP16_NEON(arm_compute::cpu::sub_same_neon<float16_t>)
+    },
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
+    {
+        "sub_same_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::U8)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::sub_same_neon<uint8_t>)
+    },
+    {
+        "sub_same_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::sub_same_neon<int16_t>)
+    },
+    {
+        "sub_same_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == data.dt3) && (data.dt1 == DataType::S32)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::sub_same_neon<int32_t>)
+    },
+    {
+        "sub_u8_s16_s16_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == DataType::U8) && (data.dt2 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::sub_u8_s16_s16_neon)
+    },
+    {
+        "sub_s16_u8_s16_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == DataType::S16) && (data.dt2 == DataType::U8)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::sub_s16_u8_s16_neon)
+    },
+    {
+        "sub_u8_u8_s16_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt3 == DataType::S16)); },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::sub_u8_u8_s16_neon)
+    },
+    {
+        "sub_qasymm8_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8)); },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::sub_qasymm8_neon)
+    },
+    {
+        "sub_qasymm8_signed_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QASYMM8_SIGNED)); },
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::sub_qasymm8_signed_neon)
+    },
+    {
+        "sub_qsymm16_neon",
+        [](const SubSelectorData & data) { return ((data.dt1 == data.dt2) && (data.dt1 == DataType::QSYMM16)); },
+        REGISTER_QSYMM16_NEON(arm_compute::cpu::sub_qsymm16_neon)
+    },
+};
+
+/** Micro-kernel selector
+ *
+ * @param[in] data Selection data passed to help pick the appropriate micro-kernel
+ *
+ * @return A matching micro-kernel else nullptr
+ */
+const SubKernel *get_implementation(DataType dt1, DataType dt2, DataType dt3)
+{
+    for(const auto &uk : available_kernels)
+    {
+        if(uk.is_selected({ dt1, dt2, dt3 }))
+        {
+            return &uk;
+        }
+    }
+    return nullptr;
+}
+
+inline Status validate_arguments(const ITensorInfo &src0, const ITensorInfo &src1, const ITensorInfo &dst, ConvertPolicy policy)
+{
+    ARM_COMPUTE_UNUSED(policy);
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&src0);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src0, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16, DataType::S16, DataType::S32, DataType::F16,
+                                                         DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16, DataType::S16, DataType::S32, DataType::F16,
+                                                         DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&dst, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16, DataType::S16, DataType::S32, DataType::F16,
+                                                         DataType::F32);
+
+    const auto *uk = get_implementation(src0.data_type(), src1.data_type(), dst.data_type());
+    ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
+
+    const TensorShape out_shape = TensorShape::broadcast_shape(src0.tensor_shape(), src1.tensor_shape());
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+        !(src0.data_type() == DataType::U8 && src1.data_type() == DataType::U8)
+        && !(src0.data_type() == DataType::QASYMM8 && src1.data_type() == DataType::QASYMM8)
+        && !(src0.data_type() == DataType::QASYMM8_SIGNED && src1.data_type() == DataType::QASYMM8_SIGNED)
+        && !(src0.data_type() == DataType::QSYMM16 && src1.data_type() == DataType::QSYMM16)
+        && !(src0.data_type() == DataType::U8 && src1.data_type() == DataType::U8)
+        && !(src0.data_type() == DataType::U8 && src1.data_type() == DataType::S16)
+        && !(src0.data_type() == DataType::S16 && src1.data_type() == DataType::U8)
+        && !(src0.data_type() == DataType::S16 && src1.data_type() == DataType::S16)
+        && !(src0.data_type() == DataType::S32 && src1.data_type() == DataType::S32)
+        && !(src0.data_type() == DataType::F32 && src1.data_type() == DataType::F32)
+        && !(src0.data_type() == DataType::F16 && src1.data_type() == DataType::F16),
+        "You called subtract with the wrong image formats");
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+        (src0.data_type() == DataType::QASYMM8_SIGNED && src1.data_type() == DataType::QASYMM8_SIGNED && policy == ConvertPolicy::WRAP)
+        || (src0.data_type() == DataType::QASYMM8 && src1.data_type() == DataType::QASYMM8 && policy == ConvertPolicy::WRAP)
+        || (src0.data_type() == DataType::QSYMM16 && src1.data_type() == DataType::QSYMM16 && policy == ConvertPolicy::WRAP),
+        "Convert policy cannot be WRAP if datatype is quantized");
+
+    // Validate in case of configured dst
+    if(dst.total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+            !(src0.data_type() == DataType::U8 && src1.data_type() == DataType::U8 && dst.data_type() == DataType::U8)
+            && !(src0.data_type() == DataType::QASYMM8 && src1.data_type() == DataType::QASYMM8 && dst.data_type() == DataType::QASYMM8)
+            && !(src0.data_type() == DataType::QASYMM8_SIGNED && src1.data_type() == DataType::QASYMM8_SIGNED && dst.data_type() == DataType::QASYMM8_SIGNED)
+            && !(src0.data_type() == DataType::QSYMM16 && src1.data_type() == DataType::QSYMM16 && dst.data_type() == DataType::QSYMM16)
+            && !(src0.data_type() == DataType::U8 && src1.data_type() == DataType::U8 && dst.data_type() == DataType::S16)
+            && !(src0.data_type() == DataType::U8 && src1.data_type() == DataType::S16 && dst.data_type() == DataType::S16)
+            && !(src0.data_type() == DataType::S16 && src1.data_type() == DataType::U8 && dst.data_type() == DataType::S16)
+            && !(src0.data_type() == DataType::S16 && src1.data_type() == DataType::S16 && dst.data_type() == DataType::S16)
+            && !(src0.data_type() == DataType::S32 && src1.data_type() == DataType::S32 && dst.data_type() == DataType::S32)
+            && !(src0.data_type() == DataType::F32 && src1.data_type() == DataType::F32 && dst.data_type() == DataType::F32)
+            && !(src0.data_type() == DataType::F16 && src1.data_type() == DataType::F16 && dst.data_type() == DataType::F16),
+            "You called subtract with the wrong image formats");
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst.tensor_shape(), 0),
+                                        "Wrong shape for dst");
+    }
+    return Status{};
+}
+} // namespace
+
+void CpuSubKernel::configure(const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst, ConvertPolicy policy)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*src0, *src1, *dst, policy));
+
+    const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*src0, *src1);
+    const TensorShape &out_shape    = broadcast_pair.first;
+    const ValidRegion &valid_region = broadcast_pair.second;
+
+    // Auto initialize dst if not initialized
+    set_shape_if_empty(*dst, out_shape);
+
+    _policy = policy;
+
+    // CpuSubKernel doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(dst->num_dimensions());
+    dst->set_valid_region(valid_region);
+    Window win = calculate_max_window(valid_region, Steps());
+
+    ICpuKernel::configure(win);
+}
+
+Status CpuSubKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, ConvertPolicy policy)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*src0, *src1, *dst, policy));
+
+    return Status{};
+}
+
+void CpuSubKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+
+    const ITensor *src0 = tensors.get_const_tensor(TensorType::ACL_SRC_0);
+    const ITensor *src1 = tensors.get_const_tensor(TensorType::ACL_SRC_1);
+    ITensor       *dst  = tensors.get_tensor(TensorType::ACL_DST);
+
+    // Dispatch kernel
+    const auto *uk = get_implementation(src0->info()->data_type(), src1->info()->data_type(), dst->info()->data_type());
+    uk->ukernel(src0, src1, dst, _policy, window);
+}
+
+const char *CpuSubKernel::name() const
+{
+    return "CpuSubKernel";
+}
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuSubKernel.h b/src/core/cpu/kernels/CpuSubKernel.h
new file mode 100644
index 0000000000..da114b6e08
--- /dev/null
+++ b/src/core/cpu/kernels/CpuSubKernel.h
@@ -0,0 +1,98 @@
+/*
+ * Copyright (c) 2016-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_CPU_SUB_KERNEL_H
+#define ARM_COMPUTE_CPU_SUB_KERNEL_H
+
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+/** Interface for the kernel to perform subtraction between two tensors */
+class CpuSubKernel : public ICpuKernel
+{
+public:
+    CpuSubKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuSubKernel);
+
+    /** Initialise the kernel's src and dst.
+     *
+     * Valid configurations (src0,src1) -> dst :
+     *
+     *   - (U8,U8)                          -> U8
+     *   - (U8,U8)                          -> S16
+     *   - (QASYMM8, QASYMM8)               -> QASYMM8
+     *   - (QASYMM8_SIGNED, QASYMM8_SIGNED) -> QASYMM8_SIGNED
+     *   - (S16,U8)                         -> S16
+     *   - (U8,S16)                         -> S16
+     *   - (S16,S16)                        -> S16
+     *   - (S32,S32)                        -> S32
+     *   - (F16,F16)                        -> F16
+     *   - (F32,F32)                        -> F32
+     *
+     * @param[in]  src0   An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/QSYMM16/S16/S32/F16/F32
+     * @param[in]  src1   An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/QSYMM16/S16/S32/F16/F32
+     * @param[out] dst    The dst tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/QSYMM16/S16/S32/F16/F32.
+     * @param[in]  policy Overflow policy. Convert policy cannot be WRAP if datatype is quantized.
+     */
+    void configure(const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst, ConvertPolicy policy);
+    /** Static function to check if given info will lead to a valid configuration of @ref CpuSubKernel
+     *
+     * Valid configurations (src0,src1) -> dst :
+     *
+     *   - (U8,U8)                          -> U8
+     *   - (U8,U8)                          -> S16
+     *   - (QASYMM8, QASYMM8)               -> QASYMM8
+     *   - (QASYMM8_SIGNED, QASYMM8_SIGNED) -> QASYMM8_SIGNED
+     *   - (S16,U8)                         -> S16
+     *   - (U8,S16)                         -> S16
+     *   - (S16,S16)                        -> S16
+     *   - (S32,S32)                        -> S32
+     *   - (F16,F16)                        -> F16
+     *   - (F32,F32)                        -> F32
+     *
+     * @param[in] src0   An input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/QSYMM16/S16/S32/F16/F32
+     * @param[in] src1   An input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/QSYMM16/S16/S32/F16/F32
+     * @param[in] dst    The dst tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/QSYMM16/S16/S32/F16/F32.
+     * @param[in] policy Policy to use to handle overflow. Convert policy cannot be WRAP if datatype is quantized.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, ConvertPolicy policy);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+
+private:
+    ConvertPolicy _policy{};
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_SUB_KERNEL_H */
diff --git a/src/core/cpu/kernels/sub/neon/integer.cpp b/src/core/cpu/kernels/sub/neon/integer.cpp
new file mode 100644
index 0000000000..bba73df1e8
--- /dev/null
+++ b/src/core/cpu/kernels/sub/neon/integer.cpp
@@ -0,0 +1,183 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace
+{
+void sub_s16_u8_s16_impl(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_swapped)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input1(src0, input1_win);
+    Iterator input2(src1, input2_win);
+    Iterator output(dst, win);
+
+    const 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());
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
+        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+        if(policy == ConvertPolicy::WRAP)
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = wrapper::vloadq(input1_ptr + x);
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                const auto res  = is_swapped ? wrapper::vsub(vin2, vin1) : wrapper::vsub(vin1, vin2);
+                wrapper::vstore(output_ptr + x, res);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto res    = is_swapped ? static_cast<int16_t>(*(input2_ptr + x)) - *(input1_ptr + x) : *(input1_ptr + x) - static_cast<int16_t>(*(input2_ptr + x));
+                *(output_ptr + x) = res;
+            }
+        }
+        else
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = wrapper::vloadq(input1_ptr + x);
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                const auto res  = is_swapped ? wrapper::vqsub(vin2, vin1) : wrapper::vqsub(vin1, vin2);
+                wrapper::vstore(output_ptr + x, res);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto res    = is_swapped ? wrapper::sub_sat(static_cast<int16_t>(*(input2_ptr + x)), *(input1_ptr + x)) : wrapper::sub_sat(*(input1_ptr + x), static_cast<int16_t>(*(input2_ptr + x)));
+                *(output_ptr + x) = res;
+            }
+        }
+    },
+    input1, input2, output);
+}
+}
+
+void sub_s16_u8_s16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    sub_s16_u8_s16_impl(src1, src0, dst, policy, window, false);
+}
+
+void sub_u8_s16_s16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    // Swap arguments
+    sub_s16_u8_s16_impl(src1, src0, dst, policy, window, true);
+}
+
+void sub_u8_u8_s16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input1(src0, input1_win);
+    Iterator input2(src1, input2_win);
+    Iterator output(dst, win);
+
+    const 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());
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
+        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+        if(policy == ConvertPolicy::WRAP)
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                wrapper::vstore(output_ptr + x, wrapper::vsub(vin1, vin2));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + x) = static_cast<int16_t>(*(input1_ptr + x)) - static_cast<int16_t>(*(input2_ptr + x));
+            }
+        }
+        else
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                wrapper::vstore(output_ptr + x, wrapper::vqsub(vin1, vin2));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + x) = wrapper::sub_sat(static_cast<int16_t>(*(input1_ptr + x)),
+                                                     static_cast<int16_t>(*(input2_ptr + x)));
+            }
+        }
+    },
+    input1, input2, output);
+}
+
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/sub/neon/list.h b/src/core/cpu/kernels/sub/neon/list.h
new file mode 100644
index 0000000000..d5685824fc
--- /dev/null
+++ b/src/core/cpu/kernels/sub/neon/list.h
@@ -0,0 +1,162 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_SUB_LIST_H
+#define SRC_CORE_NEON_KERNELS_SUB_LIST_H
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+#define DECLARE_SUB_KERNEL(func_name) \
+    void func_name(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+
+DECLARE_SUB_KERNEL(sub_qasymm8_neon);
+DECLARE_SUB_KERNEL(sub_qasymm8_signed_neon);
+DECLARE_SUB_KERNEL(sub_qsymm16_neon);
+DECLARE_SUB_KERNEL(sub_s16_u8_s16_neon);
+DECLARE_SUB_KERNEL(sub_u8_s16_s16_neon);
+DECLARE_SUB_KERNEL(sub_u8_u8_s16_neon);
+
+#undef DECLARE_SUB_KERNEL
+
+template <typename T>
+void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    /** NEON vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
+
+    bool is_sat = policy == ConvertPolicy::SATURATE;
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    constexpr int window_step_x         = 16 / sizeof(T);
+    const auto    window_start_x        = static_cast<int>(window.x().start());
+    const auto    window_end_x          = static_cast<int>(window.x().end());
+    const bool    is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+
+    Iterator input1(src0, window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()));
+    Iterator input2(src1, window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()));
+    Iterator output(dst, window);
+
+    if(is_broadcast_across_x)
+    {
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? src1 : src0;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(dst, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<T *>(output.ptr());
+
+            const T    broadcast_value     = *reinterpret_cast<const T *>(broadcast_input.ptr());
+            const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
+                auto       res             = is_sat ? wrapper::vqsub(broadcast_value_vec, non_broadcast_v) : wrapper::vsub(broadcast_value_vec, non_broadcast_v);
+                if(is_broadcast_input_2)
+                {
+                    res = wrapper::vmul(res, wrapper::vdup_n(static_cast<T>(-1), ExactTagType{}));
+                }
+                wrapper::vstore(output_ptr + x, res);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
+                auto       res             = is_sat ? wrapper::sub_sat(broadcast_value, non_broadcast_v) : broadcast_value - non_broadcast_v;
+                if(is_broadcast_input_2)
+                {
+                    res = static_cast<T>(-1) * res;
+                }
+
+                *(output_ptr + x) = res;
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<T *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto val1 = wrapper::vloadq(input1_ptr + x);
+                const auto val2 = wrapper::vloadq(input2_ptr + x);
+                const auto res  = is_sat ? wrapper::vqsub(val1, val2) : wrapper::vsub(val1, val2);
+                wrapper::vstore(output_ptr + x, res);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto val1   = *(input1_ptr + x);
+                const auto val2   = *(input2_ptr + x);
+                *(output_ptr + x) = is_sat ? wrapper::sub_sat(val1, val2) : val1 - val2;
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif // SRC_CORE_NEON_KERNELS_SUB_LIST_H
\ No newline at end of file
diff --git a/src/core/cpu/kernels/sub/neon/qasymm8.cpp b/src/core/cpu/kernels/sub/neon/qasymm8.cpp
new file mode 100644
index 0000000000..8f4cd8bdbb
--- /dev/null
+++ b/src/core/cpu/kernels/sub/neon/qasymm8.cpp
@@ -0,0 +1,230 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void sub_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    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 bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = dst->info()->quantization_info().uniform();
+
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
+    const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
+
+    if(is_broadcast_across_x)
+    {
+        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? src1 : src0;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
+        const float32x4_t             vscale1              = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
+        const float32x4_t             vscale2              = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
+        const int32x4_t               voffset1             = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
+        const int32x4_t               voffset2             = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(dst, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const uint8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<uint8_t *>(output.ptr());
+
+            const auto broadcast_value     = *reinterpret_cast<const uint8_t *>(broadcast_input.ptr());
+            const auto broadcast_value_vec = wrapper::vdup_n(static_cast<uint8_t>(broadcast_value), wrapper::traits::vector_128_tag{});
+
+            const float32x4x4_t bf =
+            {
+                {
+                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(broadcast_value_vec))))), voffset2)), vscale2),
+                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(broadcast_value_vec))))), voffset2)), vscale2),
+                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(broadcast_value_vec))))), voffset2)), vscale2),
+                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(broadcast_value_vec))))), voffset2)), vscale2),
+                }
+            };
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto a = wrapper::vloadq(non_broadcast_input_ptr + x);
+
+                const float32x4x4_t af =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
+                    }
+                };
+
+                const int32x4x4_t rf =
+                {
+                    {
+#ifdef __aarch64_
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[2], af.val[2]) : vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[3], af.val[3]) : vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#else  //__aarch64__
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[2], af.val[2]) : vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[3], af.val[3]) : vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#endif //__aarch64__
+                    }
+                };
+
+                const auto pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
+                const auto pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
+                wrapper::vstore(output_ptr + x, wrapper::vcombine(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
+                const float bfs   = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
+                *(output_ptr + x) = quantize_qasymm8(is_broadcast_input_2 ? afs - bfs : bfs - afs, dst->info()->quantization_info());
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        const float32x4_t vscale1  = vdupq_n_f32(iq1_info.scale);
+        const float32x4_t vscale2  = vdupq_n_f32(iq2_info.scale);
+        const int32x4_t   voffset1 = vdupq_n_s32(iq1_info.offset);
+        const int32x4_t   voffset2 = vdupq_n_s32(iq2_info.offset);
+
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto a = wrapper::vloadq(input1_ptr + x);
+                const auto b = wrapper::vloadq(input2_ptr + x);
+
+                const float32x4x4_t af =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
+                    }
+                };
+
+                const float32x4x4_t bf =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(b))))), voffset2)), vscale2),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(b))))), voffset2)), vscale2),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(b))))), voffset2)), vscale2),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(b))))), voffset2)), vscale2),
+                    }
+                };
+
+                const int32x4x4_t rf =
+                {
+                    {
+#ifdef __aarch64__
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#else  //__aarch64__
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#endif //__aarch64__
+                    }
+                };
+
+                const auto pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
+                const auto pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
+                wrapper::vstore(output_ptr + x, wrapper::vcombine(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale;
+                const float bfs = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale;
+
+                *(output_ptr + x) = quantize_qasymm8((afs - bfs), dst->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
+}
+
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/sub/neon/qasymm8_signed.cpp b/src/core/cpu/kernels/sub/neon/qasymm8_signed.cpp
new file mode 100644
index 0000000000..2c9e411743
--- /dev/null
+++ b/src/core/cpu/kernels/sub/neon/qasymm8_signed.cpp
@@ -0,0 +1,229 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    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 bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = dst->info()->quantization_info().uniform();
+
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
+    const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
+
+    if(is_broadcast_across_x)
+    {
+        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? src1 : src0;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
+        const float32x4_t             vscale1              = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
+        const float32x4_t             vscale2              = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
+        const int32x4_t               voffset1             = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
+        const int32x4_t               voffset2             = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(dst, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int8_t *>(output.ptr());
+
+            const auto broadcast_value     = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
+            const auto broadcast_value_vec = wrapper::vdup_n(static_cast<int8_t>(broadcast_value), wrapper::traits::vector_128_tag{});
+
+            const float32x4x4_t bf =
+            {
+                {
+                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(broadcast_value_vec))))), voffset2)), vscale2),
+                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(broadcast_value_vec))))), voffset2)), vscale2),
+                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(broadcast_value_vec))))), voffset2)), vscale2),
+                    vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(broadcast_value_vec))))), voffset2)), vscale2),
+                }
+            };
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto a = wrapper::vloadq(non_broadcast_input_ptr + x);
+
+                const float32x4x4_t af =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
+                    }
+                };
+
+                const int32x4x4_t rf =
+                {
+                    {
+#ifdef __aarch64_
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[2], af.val[2]) : vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[3], af.val[3]) : vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#else  //__aarch64__
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[2], af.val[2]) : vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[3], af.val[3]) : vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#endif //__aarch64__
+                    }
+                };
+
+                const auto pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
+                const auto pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
+                wrapper::vstore(output_ptr + x, wrapper::vcombine(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
+                const float bfs   = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
+                *(output_ptr + x) = quantize_qasymm8_signed(is_broadcast_input_2 ? afs - bfs : bfs - afs, dst->info()->quantization_info());
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        const float32x4_t vscale1  = vdupq_n_f32(iq1_info.scale);
+        const float32x4_t vscale2  = vdupq_n_f32(iq2_info.scale);
+        const int32x4_t   voffset1 = vdupq_n_s32(iq1_info.offset);
+        const int32x4_t   voffset2 = vdupq_n_s32(iq2_info.offset);
+
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto a = wrapper::vloadq(input1_ptr + x);
+                const auto b = wrapper::vloadq(input2_ptr + x);
+
+                const float32x4x4_t af =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
+                    }
+                };
+
+                const float32x4x4_t bf =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(b))))), voffset2)), vscale2),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(b))))), voffset2)), vscale2),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(b))))), voffset2)), vscale2),
+                        vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(b))))), voffset2)), vscale2),
+                    }
+                };
+
+                const int32x4x4_t rf =
+                {
+                    {
+#ifdef __aarch64__
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#else  //__aarch64__
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#endif //__aarch64__
+                    }
+                };
+
+                const auto pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
+                const auto pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
+                wrapper::vstore(output_ptr + x, wrapper::vcombine(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale;
+                const float bfs = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale;
+
+                *(output_ptr + x) = quantize_qasymm8_signed((afs - bfs), dst->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/sub/neon/qsymm16.cpp b/src/core/cpu/kernels/sub/neon/qsymm16.cpp
new file mode 100644
index 0000000000..4dfdc0e78c
--- /dev/null
+++ b/src/core/cpu/kernels/sub/neon/qsymm16.cpp
@@ -0,0 +1,201 @@
+/*
+ * 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "src/core/helpers/WindowHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void sub_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    ARM_COMPUTE_UNUSED(policy);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const 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 bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
+
+    const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = dst->info()->quantization_info().uniform();
+
+    const float32x4_t vscale1    = vdupq_n_f32(iq1_info.scale);
+    const float32x4_t vscale2    = vdupq_n_f32(iq2_info.scale);
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
+
+    if(is_broadcast_across_x)
+    {
+        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? src1 : src0;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
+        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(dst, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int16_t *>(output.ptr());
+
+            const int16_t   broadcast_value     = *reinterpret_cast<const int16_t *>(broadcast_input.ptr());
+            const int16x8_t broadcast_value_vec = vdupq_n_s16(broadcast_value);
+
+            const float32x4x2_t bf =
+            {
+                {
+                    vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(broadcast_value_vec))), vscale2),
+                    vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(broadcast_value_vec))), vscale2),
+                }
+            };
+            const float bfs = static_cast<int32_t>(broadcast_value) * broadcast_qinfo.scale;
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const int16x8_t     a = vld1q_s16(non_broadcast_input_ptr + x);
+                const float32x4x2_t af =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1),
+                    }
+                };
+
+                const int32x4x4_t rf =
+                {
+                    {
+#ifdef __aarch64__
+                        vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+#else  //__aarch64__
+                        vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+#endif //__aarch64__
+                    }
+                };
+
+                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1]));
+                vst1q_s16(output_ptr + x, pa);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x)) * non_broadcast_qinfo.scale;
+                *(output_ptr + x) = quantize_qsymm16(is_broadcast_input_2 ? (bfs - afs) : (afs - bfs), oq_info);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(src0, input1_win);
+        Iterator input2(src1, input2_win);
+        Iterator output(dst, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int16_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const int16x8_t a = vld1q_s16(input1_ptr + x);
+                const int16x8_t b = vld1q_s16(input2_ptr + x);
+
+                const float32x4x2_t af =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1),
+                    }
+                };
+
+                const float32x4x2_t bf =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(b))), vscale2),
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(b))), vscale2),
+                    }
+                };
+
+                const int32x4x2_t rf =
+                {
+                    {
+#ifdef __aarch64__
+                        vcvtnq_s32_f32(vmulq_f32(vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmulq_f32(vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+#else  //__aarch64__
+                        vcvtq_s32_f32(vmulq_f32(vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtq_s32_f32(vmulq_f32(vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+#endif //__aarch64__
+                    }
+                };
+
+                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1]));
+                vst1q_s16(output_ptr + x, pa);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = static_cast<int32_t>((*(input1_ptr + x))) * iq1_info.scale;
+                const float bfs   = static_cast<int32_t>((*(input2_ptr + x))) * iq2_info.scale;
+                *(output_ptr + x) = quantize_qsymm16((afs - bfs), dst->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file