Use CPU quantized addition kernel for quantized subtraction

Resolves : [COMPMID-5629]

Signed-off-by: Omar Al Khatib <omar.alkhatib@arm.com>
Change-Id: I061ea5bdafa3a01e66ff869d158f26a38d19e125
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8835
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

9 files changed, 540 insertions, 850 deletions

diff --git a/src/cpu/kernels/CpuSubKernel.cpp b/src/cpu/kernels/CpuSubKernel.cpp
index 37a087f115..875d613dca 100644
--- a/src/cpu/kernels/CpuSubKernel.cpp
+++ b/src/cpu/kernels/CpuSubKernel.cpp
@@ -29,14 +29,15 @@
 #include "src/core/common/Registrars.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
+#include "src/cpu/kernels/add/generic/neon/impl.h"
 #include "src/cpu/kernels/sub/neon/list.h"
 
 #if defined(ENABLE_FP32_KERNELS)
 namespace
 {
-    static constexpr size_t default_mws_N1_fp32_neon = 24385;
-    static constexpr size_t default_mws_V1_fp32_neon = 40520;
-}
+static constexpr size_t default_mws_N1_fp32_neon = 24385;
+static constexpr size_t default_mws_V1_fp32_neon = 40520;
+} // namespace
 #endif /* ENABLE_FP32_KERNELS */
 
 namespace arm_compute
@@ -47,46 +48,59 @@ namespace kernels
 {
 namespace
 {
+using CpuSubKernelDataTypeISASelectorData    = CpuAddKernelDataTypeISASelectorData;
+using CpuSubKernelDataTypeISASelectorDataPtr = CpuAddKernelDataTypeISASelectorDataPtr;
+
 static const std::vector<CpuSubKernel::SubKernel> available_kernels =
 {
     {
         "neon_fp32_sub",
-        [](const DataTypeISASelectorData & data) { return (data.dt == DataType::F32); },
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return (data.dt == DataType::F32); },
         REGISTER_FP32_NEON(arm_compute::cpu::sub_same_neon<float>)
     },
     {
         "neon_fp16_sub",
-        [](const DataTypeISASelectorData & data) { return (data.dt == DataType::F16) && data.isa.fp16; },
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return (data.dt == DataType::F16) && data.isa.fp16; },
         REGISTER_FP16_NEON(arm_compute::cpu::sub_same_neon<float16_t>)
     },
     {
         "neon_u8_sub",
-        [](const DataTypeISASelectorData & data) { return (data.dt == DataType::U8); },
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return (data.dt == DataType::U8); },
         REGISTER_INTEGER_NEON(arm_compute::cpu::sub_same_neon<uint8_t>)
     },
     {
         "neon_s16_sub",
-        [](const DataTypeISASelectorData & data) { return (data.dt == DataType::S16); },
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return (data.dt == DataType::S16); },
         REGISTER_INTEGER_NEON(arm_compute::cpu::sub_same_neon<int16_t>)
     },
     {
         "neon_s32_sub",
-        [](const DataTypeISASelectorData & data) { return (data.dt == DataType::S32); },
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return (data.dt == DataType::S32); },
         REGISTER_INTEGER_NEON(arm_compute::cpu::sub_same_neon<int32_t>)
     },
     {
+        "neon_qu8_sub_fixedpoint",
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return ((data.dt == DataType::QASYMM8) && data.can_use_fixedpoint); },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::sub_qasymm8_neon_fixedpoint)
+    },
+    {
+        "neon_qs8_sub_fixedpoint",
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return ((data.dt == DataType::QASYMM8_SIGNED) && data.can_use_fixedpoint); },
+        REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::sub_qasymm8_signed_neon_fixedpoint)
+    },
+    {
         "neon_qu8_sub",
-        [](const DataTypeISASelectorData & data) { return (data.dt == DataType::QASYMM8); },
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return (data.dt == DataType::QASYMM8); },
         REGISTER_QASYMM8_NEON(arm_compute::cpu::sub_qasymm8_neon)
     },
     {
         "neon_qs8_sub",
-        [](const DataTypeISASelectorData & data) { return (data.dt == DataType::QASYMM8_SIGNED); },
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return (data.dt == DataType::QASYMM8_SIGNED); },
         REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::sub_qasymm8_signed_neon)
     },
     {
         "neon_qs16_sub",
-        [](const DataTypeISASelectorData & data) { return (data.dt == DataType::QSYMM16); },
+        [](const CpuSubKernelDataTypeISASelectorData & data) { return (data.dt == DataType::QSYMM16); },
         REGISTER_QSYMM16_NEON(arm_compute::cpu::sub_qsymm16_neon)
     },
 };
@@ -99,7 +113,8 @@ inline Status validate_arguments(const ITensorInfo &src0, const ITensorInfo &src
                                                          DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src0, &src1);
 
-    const auto *uk = CpuSubKernel::get_implementation(DataTypeISASelectorData{ src0.data_type(), CPUInfo::get().get_isa() });
+    const auto can_use_fixedpoint = sub_q8_neon_fixedpoint_possible(&src0, &src1, &dst);
+    const auto uk                 = CpuSubKernel::get_implementation<CpuSubKernelDataTypeISASelectorData>(CpuSubKernelDataTypeISASelectorData{ src0.data_type(), CPUInfo::get().get_isa(), can_use_fixedpoint });
 
     ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
 
@@ -131,7 +146,9 @@ void CpuSubKernel::configure(const ITensorInfo *src0, const ITensorInfo *src1, I
     set_shape_if_empty(*dst, out_shape);
     set_data_type_if_unknown(*dst, src0->data_type());
 
-    const auto *uk = CpuSubKernel::get_implementation(DataTypeISASelectorData{ src0->data_type(), CPUInfo::get().get_isa() });
+    const auto can_use_fixedpoint = sub_q8_neon_fixedpoint_possible(src0, src1, dst);
+    const auto uk                 = CpuSubKernel::get_implementation<CpuSubKernelDataTypeISASelectorData>(CpuSubKernelDataTypeISASelectorData{ src0->data_type(), CPUInfo::get().get_isa(), can_use_fixedpoint });
+
     ARM_COMPUTE_ERROR_ON_NULLPTR(uk);
 
     _policy     = policy;
@@ -180,7 +197,7 @@ size_t CpuSubKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
             return std::max(static_cast<size_t>(1), mws);
         }
     }
-#else /* ENABLE_FP32_KERNELS */
+#else  /* ENABLE_FP32_KERNELS */
     ARM_COMPUTE_UNUSED(platform);
 #endif /* ENABLE_FP32_KERNELS */
     return ICPPKernel::default_mws;
diff --git a/src/cpu/kernels/CpuSubKernel.h b/src/cpu/kernels/CpuSubKernel.h
index 3d80b34279..cd209d1837 100644
--- a/src/cpu/kernels/CpuSubKernel.h
+++ b/src/cpu/kernels/CpuSubKernel.h
@@ -37,7 +37,8 @@ namespace kernels
 class CpuSubKernel : public ICpuKernel<CpuSubKernel>
 {
 private:
-    using SubKernelPtr = std::add_pointer<void(const ITensor *, const ITensor *, ITensor *, const ConvertPolicy &, const Window &)>::type;
+    using SubKernelPtr                           = std::add_pointer<void(const ITensor *, const ITensor *, ITensor *, const ConvertPolicy &, const Window &)>::type;
+    using CpuSubKernelDataTypeISASelectorDataPtr = CpuAddKernelDataTypeISASelectorDataPtr;
 
 public:
     CpuSubKernel() = default;
@@ -70,7 +71,7 @@ public:
     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;
+    void        run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
     const char *name() const override;
 
     /** Return minimum workload size of the relevant kernel
@@ -84,9 +85,9 @@ public:
 
     struct SubKernel
     {
-        const char                  *name;
-        const DataTypeISASelectorPtr is_selected;
-        SubKernelPtr                 ukernel;
+        const char                                  *name;
+        const CpuSubKernelDataTypeISASelectorDataPtr is_selected;
+        SubKernelPtr                                 ukernel;
     };
 
     static const std::vector<SubKernel> &get_available_kernels();
diff --git a/src/cpu/kernels/add/generic/neon/impl.cpp b/src/cpu/kernels/add/generic/neon/impl.cpp
index 5adb39682e..a1734d7dd6 100644
--- a/src/cpu/kernels/add/generic/neon/impl.cpp
+++ b/src/cpu/kernels/add/generic/neon/impl.cpp
@@ -64,7 +64,8 @@ void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const
         Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
         Iterator output(dst, win);
 
-        execute_window_loop(win, [&](const Coordinates &)
+        execute_window_loop(
+            win, [&](const Coordinates &)
         {
             const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr());
             const auto output_ptr              = reinterpret_cast<ScalarType *>(output.ptr());
@@ -100,7 +101,8 @@ void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const
         Iterator input2(src1, input2_win);
         Iterator output(dst, win);
 
-        execute_window_loop(win, [&](const Coordinates &)
+        execute_window_loop(
+            win, [&](const Coordinates &)
         {
             const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr());
             const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr());
@@ -128,27 +130,38 @@ void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const
     }
 }
 
+bool sub_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst)
+{
+    return add_sub_q8_neon_fixedpoint_possible(src0, src1, dst, false);
+}
+
 bool add_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst)
 {
+    return add_sub_q8_neon_fixedpoint_possible(src0, src1, dst, true);
+}
+
+bool add_sub_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, bool is_addition)
+{
     const auto iq0 = src0->quantization_info().uniform();
     const auto iq1 = src1->quantization_info().uniform();
-    const auto oq = dst->quantization_info().uniform();
+    const auto oq  = dst->quantization_info().uniform();
 
     const auto scale0 = iq0.scale / oq.scale;
     const auto scale1 = iq1.scale / oq.scale;
 
-    if(scale0 < -31.f || scale0 > 31.f || scale1 < -31.f || scale1 > 31.f)
+    if(scale0 < -15.f || scale0 > 15.f || scale1 < -15.f || scale1 > 15.f)
     {
-        // The scale factor cannot be stored as 6.10 signed fixed-point number.
+        // The scale factor cannot be stored as 5.11 signed fixed-point number.
         return false;
     }
 
     const auto offset = float(oq.offset) - scale0 * float(iq0.offset) - scale1 * float(iq1.offset);
-    const auto max_acc = (std::abs(scale0) + std::abs(scale1)) * 256.f + std::abs(offset);
 
-    if(max_acc > 2097151.f)  // 2^21 - 1
+    const auto max_acc = is_addition ? ((std::abs(scale0) + std::abs(scale1)) * 256.f + std::abs(offset)) : ((std::abs(scale0) - std::abs(scale1)) * 256.f + std::abs(offset));
+
+    if(max_acc > 1048575.f) // 2^20 - 1
     {
-        // It might not be possible to store the result as 22.10 signed fixed-point number.
+        // It might not be possible to store the result as 21.11 signed fixed-point number.
         return false;
     }
 
@@ -158,6 +171,12 @@ bool add_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo
 template <typename ScalarType>
 void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
 {
+    add_sub_q8_neon_fixedpoint<ScalarType>(src0, src1, dst, policy, window, true /*is_addition*/);
+}
+
+template <typename ScalarType>
+void add_sub_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition)
+{
     ARM_COMPUTE_UNUSED(policy);
 
     const auto in0_info = src0->info();
@@ -174,36 +193,38 @@ void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *d
     Window win = window;
     win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-    constexpr int window_step_x = 16;
-    const auto window_start_x = window.x().start();
-    const auto window_end_x = window.x().end();
-    const auto is_broadcast_across_x = in0_shape.x() != in1_shape.x();
-
-    const auto iq0_info = in0_info->quantization_info().uniform();
-    const auto iq1_info = in1_info->quantization_info().uniform();
-    const auto oq_info = dst->info()->quantization_info().uniform();
+    constexpr int window_step_x         = 16;
+    const auto    window_start_x        = window.x().start();
+    const auto    window_end_x          = window.x().end();
+    const auto    is_broadcast_across_x = in0_shape.x() != in1_shape.x();
 
+    const auto iq0_info  = in0_info->quantization_info().uniform();
+    const auto iq1_info  = in1_info->quantization_info().uniform();
+    const auto oq_info   = dst->info()->quantization_info().uniform();
     const auto in0_scale = iq0_info.scale / oq_info.scale;
-    const auto in1_scale = iq1_info.scale / oq_info.scale;
-    const auto offset = float(oq_info.offset) - in0_scale * float(iq0_info.offset) - in1_scale * float(iq1_info.offset);
+    const auto in1_scale = is_addition ? (iq1_info.scale / oq_info.scale) : (-(iq1_info.scale / oq_info.scale));
+    const auto offset    = float(oq_info.offset) - in0_scale * float(iq0_info.offset) - in1_scale * float(iq1_info.offset);
+
+    constexpr float _2pow11        = 2048;
+    const auto      in0_scale_5p11 = static_cast<int16_t>(support::cpp11::lround(in0_scale * _2pow11));
+    const auto      in1_scale_5p11 = static_cast<int16_t>(support::cpp11::lround(in1_scale * _2pow11));
+    const auto      offset_21p11   = static_cast<int32_t>(support::cpp11::lround(offset * _2pow11));
 
-    const auto in0_scale_6p10 = static_cast<int16_t>(support::cpp11::lround(in0_scale * 1024.f));
-    const auto in1_scale_6p10 = static_cast<int16_t>(support::cpp11::lround(in1_scale * 1024.f));
-    const auto offset_22p10 = static_cast<int32_t>(support::cpp11::lround(offset * 1024.f));
+    constexpr uint8_t shift_amount_remainder = 3;
 
     if(is_broadcast_across_x)
     {
         // Prefix: a = non-broadcast, b = broadcast.
 
         const auto is_broadcast_input_1 = in1_win.x().step() == 0;
-        auto a_win = is_broadcast_input_1 ? in0_win : in1_win;
-        auto b_win = is_broadcast_input_1 ? in1_win : in0_win;
-        const auto a_tensor = is_broadcast_input_1 ? src0 : src1;
-        const auto b_tensor = is_broadcast_input_1 ? src1 : src0;
+        auto       a_win                = is_broadcast_input_1 ? in0_win : in1_win;
+        auto       b_win                = is_broadcast_input_1 ? in1_win : in0_win;
+        const auto a_tensor             = is_broadcast_input_1 ? src0 : src1;
+        const auto b_tensor             = is_broadcast_input_1 ? src1 : src0;
 
-        const auto a_scale_6p10 = is_broadcast_input_1 ? in0_scale_6p10 : in1_scale_6p10;
-        const auto b_scale = is_broadcast_input_1 ? in1_scale : in0_scale;
-        const auto a_vscale_6p10 = wrapper::vdup_n(a_scale_6p10, wrapper::traits::vector_64_tag());
+        const auto a_scale_5p11  = is_broadcast_input_1 ? in0_scale_5p11 : in1_scale_5p11;
+        const auto b_scale       = is_broadcast_input_1 ? in1_scale : in0_scale;
+        const auto a_vscale_5p11 = wrapper::vdup_n(a_scale_5p11, wrapper::traits::vector_64_tag());
 
 #ifndef __aarch64__
         const auto a_scale = is_broadcast_input_1 ? in0_scale : in1_scale;
@@ -216,17 +237,18 @@ void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *d
         Iterator b_input_it(b_tensor, b_win);
         Iterator out_it(dst, win);
 
-        execute_window_loop(win, [&](const Coordinates &)
+        execute_window_loop(
+            win, [&](const Coordinates &)
         {
-            const auto a_ptr = reinterpret_cast<const ScalarType *>(a_input_it.ptr());
-            const auto b_ptr = reinterpret_cast<const ScalarType *>(b_input_it.ptr());
+            const auto a_ptr   = reinterpret_cast<const ScalarType *>(a_input_it.ptr());
+            const auto b_ptr   = reinterpret_cast<const ScalarType *>(b_input_it.ptr());
             const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr());
 
-            const auto b_val = *b_ptr;
-            const auto b_scaled = b_scale * b_val;
-            const auto b_scaled_22p10 = static_cast<int32_t>(support::cpp11::lround(b_scaled * 1024.f));
-            const auto b_scaled_offseted_22p10 = b_scaled_22p10 + offset_22p10;
-            const auto b_vscaled_offseted_22p10 = wrapper::vdup_n(b_scaled_offseted_22p10, wrapper::traits::vector_128_tag());
+            const auto b_val                    = *b_ptr;
+            const auto b_scaled                 = b_scale * b_val;
+            const auto b_scaled_21p11           = static_cast<int32_t>(support::cpp11::lround(b_scaled * _2pow11));
+            const auto b_scaled_offseted_21p11  = b_scaled_21p11 + offset_21p11;
+            const auto b_vscaled_offseted_21p11 = wrapper::vdup_n(b_scaled_offseted_21p11, wrapper::traits::vector_128_tag());
 
 #ifndef __aarch64__
             const auto b_scaled_offseted = b_scaled + offset;
@@ -245,26 +267,23 @@ void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *d
 
                 // Multiply the non-broadcast elements by the scale factor, add the scaled broadcast elements and the offset.
                 // Widen and store the result in 32-bit integer.
-                const auto vout_22p10_00 = wrapper::vmlal(b_vscaled_offseted_22p10, wrapper::vgetlow(a_vin_16p0_0), a_vscale_6p10);
-                const auto vout_22p10_01 = wrapper::vmlal(b_vscaled_offseted_22p10, wrapper::vgethigh(a_vin_16p0_0), a_vscale_6p10);
-                const auto vout_22p10_10 = wrapper::vmlal(b_vscaled_offseted_22p10, wrapper::vgetlow(a_vin_16p0_1), a_vscale_6p10);
-                const auto vout_22p10_11 = wrapper::vmlal(b_vscaled_offseted_22p10, wrapper::vgethigh(a_vin_16p0_1), a_vscale_6p10);
+                const auto vout_21p11_00 = wrapper::vmlal(b_vscaled_offseted_21p11, wrapper::vgetlow(a_vin_16p0_0), a_vscale_5p11);
+                const auto vout_21p11_01 = wrapper::vmlal(b_vscaled_offseted_21p11, wrapper::vgethigh(a_vin_16p0_0), a_vscale_5p11);
+                const auto vout_21p11_10 = wrapper::vmlal(b_vscaled_offseted_21p11, wrapper::vgetlow(a_vin_16p0_1), a_vscale_5p11);
+                const auto vout_21p11_11 = wrapper::vmlal(b_vscaled_offseted_21p11, wrapper::vgethigh(a_vin_16p0_1), a_vscale_5p11);
 
-                // Remove 2 bits of the fractional part, round, narrow to 16-bit and saturate the result.
+                // Remove 3 bits of the fractional part, round, narrow to 16-bit and saturate the result.
                 const auto vout_8p8_0 = wrapper::vcombine(
-                    wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_00),
-                    wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_01)
-                );
+                                            wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(vout_21p11_00),
+                                            wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(vout_21p11_01));
                 const auto vout_8p8_1 = wrapper::vcombine(
-                    wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_10),
-                    wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_11)
-                );
+                                            wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(vout_21p11_10),
+                                            wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(vout_21p11_11));
 
                 // Remove 8 bits of the fractional part, round, narrow to 8-bit and saturate the result.
                 const auto vout_8p0 = wrapper::vcombine(
-                    wrapper::vqrshrn<8>(vout_8p8_0),
-                    wrapper::vqrshrn<8>(vout_8p8_1)
-                );
+                                          wrapper::vqrshrn<8>(vout_8p8_0),
+                                          wrapper::vqrshrn<8>(vout_8p8_1));
 
                 // Store the result.
                 wrapper::vstore(out_ptr + x, vout_8p0);
@@ -274,8 +293,8 @@ void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *d
             for(; x < window_end_x; ++x)
             {
 #ifdef __aarch64__
-                out_ptr[x] = wrapper::vqrshrn<8>(wrapper::vqrshrn_ex<2, ScalarType>(int32_t(a_ptr[x]) * a_scale_6p10 + b_scaled_offseted_22p10));
-#else // __aarch64__
+                out_ptr[x] = wrapper::vqrshrn<8>(wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(int32_t(a_ptr[x]) * a_scale_5p11 + b_scaled_offseted_21p11));
+#else  // __aarch64__
                 out_ptr[x] = utility::clamp<int, ScalarType>(support::cpp11::lround(float(a_ptr[x]) * a_scale + b_scaled_offseted));
 #endif // __aarch64__
             }
@@ -284,9 +303,9 @@ void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *d
     }
     else
     {
-        const auto vscale0_6p10 = wrapper::vdup_n(in0_scale_6p10, wrapper::traits::vector_64_tag());
-        const auto vscale1_6p10 = wrapper::vdup_n(in1_scale_6p10, wrapper::traits::vector_64_tag());
-        const auto voffset_22p10 = wrapper::vdup_n(offset_22p10, wrapper::traits::vector_128_tag());
+        const auto vscale0_5p11  = wrapper::vdup_n(in0_scale_5p11, wrapper::traits::vector_64_tag());
+        const auto vscale1_5p11  = wrapper::vdup_n(in1_scale_5p11, wrapper::traits::vector_64_tag());
+        const auto voffset_21p11 = wrapper::vdup_n(offset_21p11, wrapper::traits::vector_128_tag());
 
         // Clear the x dimension on the execution window as we process the whole row each iteration.
         in0_win.set(Window::DimX, Window::Dimension(0, 1, 1));
@@ -296,7 +315,8 @@ void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *d
         Iterator in1_it(src1, in1_win);
         Iterator out_it(dst, win);
 
-        execute_window_loop(win, [&](const Coordinates &)
+        execute_window_loop(
+            win, [&](const Coordinates &)
         {
             const auto in0_ptr = reinterpret_cast<const ScalarType *>(in0_it.ptr());
             const auto in1_ptr = reinterpret_cast<const ScalarType *>(in1_it.ptr());
@@ -318,31 +338,28 @@ void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *d
 
                 // Multiply the input elements by the scale factor and add the offset.
                 // Widen and store the result in 32-bit integer.
-                const auto vscaled0_offseted_22p10_00 = wrapper::vmlal(voffset_22p10, wrapper::vgetlow(vin0_16p0_0), vscale0_6p10);
-                const auto vscaled0_offseted_22p10_01 = wrapper::vmlal(voffset_22p10, wrapper::vgethigh(vin0_16p0_0), vscale0_6p10);
-                const auto vscaled0_offseted_22p10_10 = wrapper::vmlal(voffset_22p10, wrapper::vgetlow(vin0_16p0_1), vscale0_6p10);
-                const auto vscaled0_offseted_22p10_11 = wrapper::vmlal(voffset_22p10, wrapper::vgethigh(vin0_16p0_1), vscale0_6p10);
+                const auto vscaled0_offseted_21p11_00 = wrapper::vmlal(voffset_21p11, wrapper::vgetlow(vin0_16p0_0), vscale0_5p11);
+                const auto vscaled0_offseted_21p11_01 = wrapper::vmlal(voffset_21p11, wrapper::vgethigh(vin0_16p0_0), vscale0_5p11);
+                const auto vscaled0_offseted_21p11_10 = wrapper::vmlal(voffset_21p11, wrapper::vgetlow(vin0_16p0_1), vscale0_5p11);
+                const auto vscaled0_offseted_21p11_11 = wrapper::vmlal(voffset_21p11, wrapper::vgethigh(vin0_16p0_1), vscale0_5p11);
 
-                const auto vout_22p10_00 = wrapper::vmlal(vscaled0_offseted_22p10_00, wrapper::vgetlow(vin1_16p0_0), vscale1_6p10);
-                const auto vout_22p10_01 = wrapper::vmlal(vscaled0_offseted_22p10_01, wrapper::vgethigh(vin1_16p0_0), vscale1_6p10);
-                const auto vout_22p10_10 = wrapper::vmlal(vscaled0_offseted_22p10_10, wrapper::vgetlow(vin1_16p0_1), vscale1_6p10);
-                const auto vout_22p10_11 = wrapper::vmlal(vscaled0_offseted_22p10_11, wrapper::vgethigh(vin1_16p0_1), vscale1_6p10);
+                const auto vout_21p11_00 = wrapper::vmlal(vscaled0_offseted_21p11_00, wrapper::vgetlow(vin1_16p0_0), vscale1_5p11);
+                const auto vout_21p11_01 = wrapper::vmlal(vscaled0_offseted_21p11_01, wrapper::vgethigh(vin1_16p0_0), vscale1_5p11);
+                const auto vout_21p11_10 = wrapper::vmlal(vscaled0_offseted_21p11_10, wrapper::vgetlow(vin1_16p0_1), vscale1_5p11);
+                const auto vout_21p11_11 = wrapper::vmlal(vscaled0_offseted_21p11_11, wrapper::vgethigh(vin1_16p0_1), vscale1_5p11);
 
-                // Remove 2 bits of the fractional part, round, narrow to 16-bit and saturate the result.
+                // Remove 3 bits of the fractional part, round, narrow to 16-bit and saturate the result.
                 const auto vout_8p8_0 = wrapper::vcombine(
-                    wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_00),
-                    wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_01)
-                );
+                                            wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(vout_21p11_00),
+                                            wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(vout_21p11_01));
                 const auto vout_8p8_1 = wrapper::vcombine(
-                    wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_10),
-                    wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_11)
-                );
+                                            wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(vout_21p11_10),
+                                            wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(vout_21p11_11));
 
                 // Remove 8 bits of the fractional part, round, narrow to 8-bit and saturate the result.
                 const auto vout_8p0 = wrapper::vcombine(
-                    wrapper::vqrshrn<8>(vout_8p8_0),
-                    wrapper::vqrshrn<8>(vout_8p8_1)
-                );
+                                          wrapper::vqrshrn<8>(vout_8p8_0),
+                                          wrapper::vqrshrn<8>(vout_8p8_1));
 
                 // Store the result.
                 wrapper::vstore(out_ptr + x, vout_8p0);
@@ -352,8 +369,8 @@ void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *d
             for(; x < window_end_x; ++x)
             {
 #ifdef __aarch64__
-                out_ptr[x] = wrapper::vqrshrn<8>(wrapper::vqrshrn_ex<2, ScalarType>(int32_t(in0_ptr[x]) * in0_scale_6p10 + int32_t(in1_ptr[x]) * in1_scale_6p10 + offset_22p10));
-#else // __aarch64__
+                out_ptr[x] = wrapper::vqrshrn<8>(wrapper::vqrshrn_ex<shift_amount_remainder, ScalarType>(int32_t(in0_ptr[x]) * in0_scale_5p11 + int32_t(in1_ptr[x]) * in1_scale_5p11 + offset_21p11));
+#else  // __aarch64__
                 out_ptr[x] = utility::clamp<int, ScalarType>(support::cpp11::lround(float(in0_ptr[x]) * in0_scale + float(in1_ptr[x]) * in1_scale + offset));
 #endif // __aarch64__
             }
@@ -362,6 +379,372 @@ void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *d
     }
 }
 
+void add_sub_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition)
+{
+    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));
+
+    constexpr int window_step_x         = 16;
+    const auto    window_start_x        = static_cast<int>(window.x().start());
+    const auto    window_end_x          = static_cast<int>(window.x().end());
+    const 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 auto scale1 = iq1_info.scale / oq_info.scale;
+    const auto scale2 = is_addition ? (iq2_info.scale / oq_info.scale) : (-(iq2_info.scale / oq_info.scale));
+    const auto offset = float(oq_info.offset) - scale1 * float(iq1_info.offset) - scale2 * float(iq2_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 auto af_scale = is_broadcast_input_2 ? scale1 : scale2;
+        const auto bf_scale = is_broadcast_input_2 ? scale2 : scale1;
+        const auto vscale1  = vdupq_n_f32(af_scale);
+
+        // 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 = non_broadcast_input.ptr();
+            const auto output_ptr              = output.ptr();
+
+            const auto broadcast_value = *broadcast_input.ptr();
+            const auto bf              = vdupq_n_f32(float(broadcast_value) * scale2 + offset);
+            const auto bfs             = float(broadcast_value) * bf_scale + offset;
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const uint8x16_t a = vld1q_u8(non_broadcast_input_ptr + x);
+
+                const auto a_u16_0 = vmovl_u8(vget_low_u8(a));
+                const auto a_u16_1 = vmovl_u8(vget_high_u8(a));
+
+                const auto af_0 = vmlaq_f32(bf, vcvtq_f32_u32(vmovl_u16(vget_low_u16(a_u16_0))), vscale1);
+                const auto af_1 = vmlaq_f32(bf, vcvtq_f32_u32(vmovl_u16(vget_high_u16(a_u16_0))), vscale1);
+                const auto af_2 = vmlaq_f32(bf, vcvtq_f32_u32(vmovl_u16(vget_low_u16(a_u16_1))), vscale1);
+                const auto af_3 = vmlaq_f32(bf, vcvtq_f32_u32(vmovl_u16(vget_high_u16(a_u16_1))), vscale1);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+                int32x4_t rf_2{};
+                int32x4_t rf_3{};
+
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(af_0);
+                rf_1 = vcvtnq_s32_f32(af_1);
+                rf_2 = vcvtnq_s32_f32(af_2);
+                rf_3 = vcvtnq_s32_f32(af_3);
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(af_0);
+                rf_1 = vcvtq_s32_f32(af_1);
+                rf_2 = vcvtq_s32_f32(af_2);
+                rf_3 = vcvtq_s32_f32(af_3);
+#endif //__aarch64__
+
+                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
+                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
+                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto result = float(non_broadcast_input_ptr[x]) * af_scale + bfs;
+#ifdef __aarch64__
+                output_ptr[x] = utility::clamp<int, uint8_t>(support::cpp11::lround(result));
+#else  // __aarch64__
+                output_ptr[x] = utility::clamp<int, uint8_t>(support::cpp11::trunc(result));
+#endif // __aarch64__
+            }
+        },
+        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);
+
+        const auto vscale1 = vdupq_n_f32(scale1);
+        const auto vscale2 = vdupq_n_f32(scale2);
+        const auto voffset = vdupq_n_f32(offset);
+
+        execute_window_loop(
+            win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = input1.ptr();
+            const auto input2_ptr = input2.ptr();
+            const auto output_ptr = 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 uint8x16_t a = vld1q_u8(input1_ptr + x);
+                const uint8x16_t b = vld1q_u8(input2_ptr + x);
+
+                const auto a_u16_0 = vmovl_u8(vget_low_u8(a));
+                const auto a_u16_1 = vmovl_u8(vget_high_u8(a));
+                const auto b_u16_0 = vmovl_u8(vget_low_u8(b));
+                const auto b_u16_1 = vmovl_u8(vget_high_u8(b));
+
+                const auto af_0 = vmlaq_f32(voffset, vcvtq_f32_u32(vmovl_u16(vget_low_u16(a_u16_0))), vscale1);
+                const auto af_1 = vmlaq_f32(voffset, vcvtq_f32_u32(vmovl_u16(vget_high_u16(a_u16_0))), vscale1);
+                const auto af_2 = vmlaq_f32(voffset, vcvtq_f32_u32(vmovl_u16(vget_low_u16(a_u16_1))), vscale1);
+                const auto af_3 = vmlaq_f32(voffset, vcvtq_f32_u32(vmovl_u16(vget_high_u16(a_u16_1))), vscale1);
+
+                const auto bf_0 = vmlaq_f32(af_0, vcvtq_f32_u32(vmovl_u16(vget_low_u16(b_u16_0))), vscale2);
+                const auto bf_1 = vmlaq_f32(af_1, vcvtq_f32_u32(vmovl_u16(vget_high_u16(b_u16_0))), vscale2);
+                const auto bf_2 = vmlaq_f32(af_2, vcvtq_f32_u32(vmovl_u16(vget_low_u16(b_u16_1))), vscale2);
+                const auto bf_3 = vmlaq_f32(af_3, vcvtq_f32_u32(vmovl_u16(vget_high_u16(b_u16_1))), vscale2);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+                int32x4_t rf_2{};
+                int32x4_t rf_3{};
+
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(bf_0);
+                rf_1 = vcvtnq_s32_f32(bf_1);
+                rf_2 = vcvtnq_s32_f32(bf_2);
+                rf_3 = vcvtnq_s32_f32(bf_3);
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(bf_0);
+                rf_1 = vcvtq_s32_f32(bf_1);
+                rf_2 = vcvtq_s32_f32(bf_2);
+                rf_3 = vcvtq_s32_f32(bf_3);
+#endif //__aarch64__
+
+                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
+                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
+                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto result = float(input1_ptr[x]) * scale1 + float(input2_ptr[x]) * scale2 + offset;
+#ifdef __aarch64__
+                output_ptr[x] = utility::clamp<int, uint8_t>(support::cpp11::lround(result));
+#else  // __aarch64__
+                output_ptr[x] = utility::clamp<int, uint8_t>(support::cpp11::trunc(result));
+#endif // __aarch64__
+            }
+        },
+        input1, input2, output);
+    }
+}
+
+void add_sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition)
+{
+    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));
+
+    constexpr int window_step_x         = 16;
+    const auto    window_start_x        = static_cast<int>(window.x().start());
+    const auto    window_end_x          = static_cast<int>(window.x().end());
+    const 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 auto scale1 = iq1_info.scale / oq_info.scale;
+    const auto scale2 = is_addition ? (iq2_info.scale / oq_info.scale) : (-(iq2_info.scale / oq_info.scale));
+    const auto offset = float(oq_info.offset) - scale1 * float(iq1_info.offset) - scale2 * float(iq2_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 auto af_scale = is_broadcast_input_2 ? scale1 : scale2;
+        const auto bf_scale = is_broadcast_input_2 ? scale2 : scale1;
+        const auto vscale1  = vdupq_n_f32(af_scale);
+
+        // 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 bf              = vdupq_n_f32(float(broadcast_value) * scale2 + offset);
+            const auto bfs             = float(broadcast_value) * bf_scale + offset;
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const int8x16_t a = vld1q_s8(non_broadcast_input_ptr + x);
+
+                const auto a_s16_0 = vmovl_s8(vget_low_s8(a));
+                const auto a_s16_1 = vmovl_s8(vget_high_s8(a));
+
+                const auto af_0 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_0))), vscale1);
+                const auto af_1 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_0))), vscale1);
+                const auto af_2 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_1))), vscale1);
+                const auto af_3 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_1))), vscale1);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+                int32x4_t rf_2{};
+                int32x4_t rf_3{};
+
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(af_0);
+                rf_1 = vcvtnq_s32_f32(af_1);
+                rf_2 = vcvtnq_s32_f32(af_2);
+                rf_3 = vcvtnq_s32_f32(af_3);
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(af_0);
+                rf_1 = vcvtq_s32_f32(af_1);
+                rf_2 = vcvtq_s32_f32(af_2);
+                rf_3 = vcvtq_s32_f32(af_3);
+#endif //__aarch64__
+
+                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
+                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
+                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto result = float(non_broadcast_input_ptr[x]) * af_scale + bfs;
+#ifdef __aarch64__
+                output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::lround(result));
+#else  // __aarch64__
+                output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::trunc(result));
+#endif // __aarch64__
+            }
+        },
+        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);
+
+        const auto vscale1 = vdupq_n_f32(scale1);
+        const auto vscale2 = vdupq_n_f32(scale2);
+        const auto voffset = vdupq_n_f32(offset);
+
+        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 int8x16_t a = vld1q_s8(input1_ptr + x);
+                const int8x16_t b = vld1q_s8(input2_ptr + x);
+
+                const auto a_s16_0 = vmovl_s8(vget_low_s8(a));
+                const auto a_s16_1 = vmovl_s8(vget_high_s8(a));
+                const auto b_s16_0 = vmovl_s8(vget_low_s8(b));
+                const auto b_s16_1 = vmovl_s8(vget_high_s8(b));
+
+                const auto af_0 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_0))), vscale1);
+                const auto af_1 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_0))), vscale1);
+                const auto af_2 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_1))), vscale1);
+                const auto af_3 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_1))), vscale1);
+
+                const auto bf_0 = vmlaq_f32(af_0, vcvtq_f32_s32(vmovl_s16(vget_low_s16(b_s16_0))), vscale2);
+                const auto bf_1 = vmlaq_f32(af_1, vcvtq_f32_s32(vmovl_s16(vget_high_s16(b_s16_0))), vscale2);
+                const auto bf_2 = vmlaq_f32(af_2, vcvtq_f32_s32(vmovl_s16(vget_low_s16(b_s16_1))), vscale2);
+                const auto bf_3 = vmlaq_f32(af_3, vcvtq_f32_s32(vmovl_s16(vget_high_s16(b_s16_1))), vscale2);
+
+                int32x4_t rf_0{};
+                int32x4_t rf_1{};
+                int32x4_t rf_2{};
+                int32x4_t rf_3{};
+
+#ifdef __aarch64__
+                rf_0 = vcvtnq_s32_f32(bf_0);
+                rf_1 = vcvtnq_s32_f32(bf_1);
+                rf_2 = vcvtnq_s32_f32(bf_2);
+                rf_3 = vcvtnq_s32_f32(bf_3);
+#else  //__aarch64__
+                rf_0 = vcvtq_s32_f32(bf_0);
+                rf_1 = vcvtq_s32_f32(bf_1);
+                rf_2 = vcvtq_s32_f32(bf_2);
+                rf_3 = vcvtq_s32_f32(bf_3);
+#endif //__aarch64__
+
+                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
+                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
+                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto result = float(input1_ptr[x]) * scale1 + float(input2_ptr[x]) * scale2 + offset;
+#ifdef __aarch64__
+                output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::lround(result));
+#else  // __aarch64__
+                output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::trunc(result));
+#endif // __aarch64__
+            }
+        },
+        input1, input2, output);
+    }
+}
+
 template void add_same_neon<float>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
 template void add_same_neon<uint8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
 template void add_same_neon<int32_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
@@ -374,5 +757,11 @@ template void add_same_neon<float16_t>(const ITensor *src0, const ITensor *src1,
 template void add_q8_neon_fixedpoint<int8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
 template void add_q8_neon_fixedpoint<uint8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
 
+template void add_sub_q8_neon_fixedpoint<int8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition);
+template void add_sub_q8_neon_fixedpoint<uint8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition);
+
+void add_sub_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition);
+void add_sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition);
+
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/cpu/kernels/add/generic/neon/impl.h b/src/cpu/kernels/add/generic/neon/impl.h
index 91f347ff9c..d544ef5728 100644
--- a/src/cpu/kernels/add/generic/neon/impl.h
+++ b/src/cpu/kernels/add/generic/neon/impl.h
@@ -35,8 +35,19 @@ void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const
 
 bool add_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst);
 
+bool sub_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst);
+
+bool add_sub_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, bool is_addition);
+
+void add_sub_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition);
+
+void add_sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition);
+
 template <typename ScalarType>
 void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
+
+template <typename ScalarType>
+void add_sub_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window, bool is_addition);
 } // namespace cpu
 } // namespace arm_compute
 #endif // SRC_CORE_NEON_KERNELS_ADD_IMPL_H
\ No newline at end of file
diff --git a/src/cpu/kernels/add/generic/neon/qasymm8.cpp b/src/cpu/kernels/add/generic/neon/qasymm8.cpp
index d8b4bca292..69cca956c8 100644
--- a/src/cpu/kernels/add/generic/neon/qasymm8.cpp
+++ b/src/cpu/kernels/add/generic/neon/qasymm8.cpp
@@ -21,12 +21,8 @@
  * 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"
 #include "src/cpu/kernels/add/generic/neon/impl.h"
 
 namespace arm_compute
@@ -35,183 +31,7 @@ namespace cpu
 {
 void add_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));
-
-    constexpr int window_step_x      = 16;
-    const auto window_start_x        = static_cast<int>(window.x().start());
-    const auto window_end_x          = static_cast<int>(window.x().end());
-    const 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 auto scale1 = iq1_info.scale / oq_info.scale;
-    const auto scale2 = iq2_info.scale / oq_info.scale;
-    const auto offset = float(oq_info.offset) - scale1 * float(iq1_info.offset) - scale2 * float(iq2_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 auto af_scale = is_broadcast_input_2 ? scale1 : scale2;
-        const auto bf_scale = is_broadcast_input_2 ? scale2 : scale1;
-        const auto vscale1  = vdupq_n_f32(af_scale);
-
-        // 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 = non_broadcast_input.ptr();
-            const auto output_ptr              = output.ptr();
-
-            const auto broadcast_value = *broadcast_input.ptr();
-            const auto bf = vdupq_n_f32(float(broadcast_value) * scale2 + offset);
-            const auto bfs = float(broadcast_value) * bf_scale + offset;
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const uint8x16_t a    = vld1q_u8(non_broadcast_input_ptr + x);
-
-                const auto a_u16_0 = vmovl_u8(vget_low_u8(a));
-                const auto a_u16_1 = vmovl_u8(vget_high_u8(a));
-
-                const auto af_0 = vmlaq_f32(bf, vcvtq_f32_u32(vmovl_u16(vget_low_u16(a_u16_0))), vscale1);
-                const auto af_1 = vmlaq_f32(bf, vcvtq_f32_u32(vmovl_u16(vget_high_u16(a_u16_0))), vscale1);
-                const auto af_2 = vmlaq_f32(bf, vcvtq_f32_u32(vmovl_u16(vget_low_u16(a_u16_1))), vscale1);
-                const auto af_3 = vmlaq_f32(bf, vcvtq_f32_u32(vmovl_u16(vget_high_u16(a_u16_1))), vscale1);
-
-                int32x4_t rf_0{};
-                int32x4_t rf_1{};
-                int32x4_t rf_2{};
-                int32x4_t rf_3{};
-
-#ifdef __aarch64__
-                rf_0 = vcvtnq_s32_f32(af_0);
-                rf_1 = vcvtnq_s32_f32(af_1);
-                rf_2 = vcvtnq_s32_f32(af_2);
-                rf_3 = vcvtnq_s32_f32(af_3);
-#else  //__aarch64__
-                rf_0 = vcvtq_s32_f32(af_0);
-                rf_1 = vcvtq_s32_f32(af_1);
-                rf_2 = vcvtq_s32_f32(af_2);
-                rf_3 = vcvtq_s32_f32(af_3);
-#endif //__aarch64__
-
-                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
-                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
-                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const auto result = float(non_broadcast_input_ptr[x]) * af_scale + bfs;
-#ifdef __aarch64__
-                output_ptr[x] = utility::clamp<int, uint8_t>(support::cpp11::lround(result));
-#else  // __aarch64__
-                output_ptr[x] = utility::clamp<int, uint8_t>(support::cpp11::trunc(result));
-#endif  // __aarch64__
-            }
-        },
-        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);
-
-        const auto vscale1 = vdupq_n_f32(scale1);
-        const auto vscale2 = vdupq_n_f32(scale2);
-        const auto voffset = vdupq_n_f32(offset);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto input1_ptr = input1.ptr();
-            const auto input2_ptr = input2.ptr();
-            const auto output_ptr = 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 uint8x16_t a = vld1q_u8(input1_ptr + x);
-                const uint8x16_t b = vld1q_u8(input2_ptr + x);
-
-                const auto a_u16_0 = vmovl_u8(vget_low_u8(a));
-                const auto a_u16_1 = vmovl_u8(vget_high_u8(a));
-                const auto b_u16_0 = vmovl_u8(vget_low_u8(b));
-                const auto b_u16_1 = vmovl_u8(vget_high_u8(b));
-
-                const auto af_0 = vmlaq_f32(voffset, vcvtq_f32_u32(vmovl_u16(vget_low_u16(a_u16_0))), vscale1);
-                const auto af_1 = vmlaq_f32(voffset, vcvtq_f32_u32(vmovl_u16(vget_high_u16(a_u16_0))), vscale1);
-                const auto af_2 = vmlaq_f32(voffset, vcvtq_f32_u32(vmovl_u16(vget_low_u16(a_u16_1))), vscale1);
-                const auto af_3 = vmlaq_f32(voffset, vcvtq_f32_u32(vmovl_u16(vget_high_u16(a_u16_1))), vscale1);
-
-                const auto bf_0 = vmlaq_f32(af_0, vcvtq_f32_u32(vmovl_u16(vget_low_u16(b_u16_0))), vscale2);
-                const auto bf_1 = vmlaq_f32(af_1, vcvtq_f32_u32(vmovl_u16(vget_high_u16(b_u16_0))), vscale2);
-                const auto bf_2 = vmlaq_f32(af_2, vcvtq_f32_u32(vmovl_u16(vget_low_u16(b_u16_1))), vscale2);
-                const auto bf_3 = vmlaq_f32(af_3, vcvtq_f32_u32(vmovl_u16(vget_high_u16(b_u16_1))), vscale2);
-
-                int32x4_t rf_0{};
-                int32x4_t rf_1{};
-                int32x4_t rf_2{};
-                int32x4_t rf_3{};
-
-#ifdef __aarch64__
-                rf_0 = vcvtnq_s32_f32(bf_0);
-                rf_1 = vcvtnq_s32_f32(bf_1);
-                rf_2 = vcvtnq_s32_f32(bf_2);
-                rf_3 = vcvtnq_s32_f32(bf_3);
-#else  //__aarch64__
-                rf_0 = vcvtq_s32_f32(bf_0);
-                rf_1 = vcvtq_s32_f32(bf_1);
-                rf_2 = vcvtq_s32_f32(bf_2);
-                rf_3 = vcvtq_s32_f32(bf_3);
-#endif //__aarch64__
-
-                const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
-                const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
-                vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const auto result = float(input1_ptr[x]) * scale1 + float(input2_ptr[x]) * scale2 + offset;
-#ifdef __aarch64__
-                output_ptr[x] = utility::clamp<int, uint8_t>(support::cpp11::lround(result));
-#else  // __aarch64__
-                output_ptr[x] = utility::clamp<int, uint8_t>(support::cpp11::trunc(result));
-#endif  // __aarch64__
-            }
-        },
-        input1, input2, output);
-    }
+    add_sub_qasymm8_neon(src0, src1, dst, policy, window, true /*is_addition*/);
 }
 } // namespace cpu
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp b/src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp
index a285e483ed..dfdf8fe85b 100644
--- a/src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp
+++ b/src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp
@@ -21,12 +21,8 @@
  * 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"
 #include "src/cpu/kernels/add/generic/neon/impl.h"
 
 namespace arm_compute
@@ -35,183 +31,7 @@ namespace cpu
 {
 void add_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));
-
-    constexpr int window_step_x      = 16;
-    const auto window_start_x        = static_cast<int>(window.x().start());
-    const auto window_end_x          = static_cast<int>(window.x().end());
-    const 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 auto scale1 = iq1_info.scale / oq_info.scale;
-    const auto scale2 = iq2_info.scale / oq_info.scale;
-    const auto offset = float(oq_info.offset) - scale1 * float(iq1_info.offset) - scale2 * float(iq2_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 auto af_scale = is_broadcast_input_2 ? scale1 : scale2;
-        const auto bf_scale = is_broadcast_input_2 ? scale2 : scale1;
-        const auto vscale1  = vdupq_n_f32(af_scale);
-
-        // 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 bf = vdupq_n_f32(float(broadcast_value) * scale2 + offset);
-            const auto bfs = float(broadcast_value) * bf_scale + offset;
-
-            // Compute S elements per iteration
-            int x = window_start_x;
-            for(; x <= (window_end_x - window_step_x); x += window_step_x)
-            {
-                const int8x16_t a = vld1q_s8(non_broadcast_input_ptr + x);
-
-                const auto a_s16_0 = vmovl_s8(vget_low_s8(a));
-                const auto a_s16_1 = vmovl_s8(vget_high_s8(a));
-
-                const auto af_0 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_0))), vscale1);
-                const auto af_1 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_0))), vscale1);
-                const auto af_2 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_1))), vscale1);
-                const auto af_3 = vmlaq_f32(bf, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_1))), vscale1);
-
-                int32x4_t rf_0{};
-                int32x4_t rf_1{};
-                int32x4_t rf_2{};
-                int32x4_t rf_3{};
-
-#ifdef __aarch64__
-                rf_0 = vcvtnq_s32_f32(af_0);
-                rf_1 = vcvtnq_s32_f32(af_1);
-                rf_2 = vcvtnq_s32_f32(af_2);
-                rf_3 = vcvtnq_s32_f32(af_3);
-#else  //__aarch64__
-                rf_0 = vcvtq_s32_f32(af_0);
-                rf_1 = vcvtq_s32_f32(af_1);
-                rf_2 = vcvtq_s32_f32(af_2);
-                rf_3 = vcvtq_s32_f32(af_3);
-#endif //__aarch64__
-
-                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
-                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
-                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const auto result = float(non_broadcast_input_ptr[x]) * af_scale + bfs;
-#ifdef __aarch64__
-                output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::lround(result));
-#else  // __aarch64__
-                output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::trunc(result));
-#endif  // __aarch64__
-            }
-        },
-        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);
-
-        const auto vscale1 = vdupq_n_f32(scale1);
-        const auto vscale2 = vdupq_n_f32(scale2);
-        const auto voffset = vdupq_n_f32(offset);
-
-        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 int8x16_t a = vld1q_s8(input1_ptr + x);
-                const int8x16_t b = vld1q_s8(input2_ptr + x);
-
-                const auto a_s16_0 = vmovl_s8(vget_low_s8(a));
-                const auto a_s16_1 = vmovl_s8(vget_high_s8(a));
-                const auto b_s16_0 = vmovl_s8(vget_low_s8(b));
-                const auto b_s16_1 = vmovl_s8(vget_high_s8(b));
-
-                const auto af_0 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_0))), vscale1);
-                const auto af_1 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_0))), vscale1);
-                const auto af_2 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_low_s16(a_s16_1))), vscale1);
-                const auto af_3 = vmlaq_f32(voffset, vcvtq_f32_s32(vmovl_s16(vget_high_s16(a_s16_1))), vscale1);
-
-                const auto bf_0 = vmlaq_f32(af_0, vcvtq_f32_s32(vmovl_s16(vget_low_s16(b_s16_0))), vscale2);
-                const auto bf_1 = vmlaq_f32(af_1, vcvtq_f32_s32(vmovl_s16(vget_high_s16(b_s16_0))), vscale2);
-                const auto bf_2 = vmlaq_f32(af_2, vcvtq_f32_s32(vmovl_s16(vget_low_s16(b_s16_1))), vscale2);
-                const auto bf_3 = vmlaq_f32(af_3, vcvtq_f32_s32(vmovl_s16(vget_high_s16(b_s16_1))), vscale2);
-
-                int32x4_t rf_0{};
-                int32x4_t rf_1{};
-                int32x4_t rf_2{};
-                int32x4_t rf_3{};
-
-#ifdef __aarch64__
-                rf_0 = vcvtnq_s32_f32(bf_0);
-                rf_1 = vcvtnq_s32_f32(bf_1);
-                rf_2 = vcvtnq_s32_f32(bf_2);
-                rf_3 = vcvtnq_s32_f32(bf_3);
-#else  //__aarch64__
-                rf_0 = vcvtq_s32_f32(bf_0);
-                rf_1 = vcvtq_s32_f32(bf_1);
-                rf_2 = vcvtq_s32_f32(bf_2);
-                rf_3 = vcvtq_s32_f32(bf_3);
-#endif //__aarch64__
-
-                const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
-                const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
-                vst1q_s8(output_ptr + x, vcombine_s8(pa, pb));
-            }
-
-            // Compute left-over elements
-            for(; x < window_end_x; ++x)
-            {
-                const auto result = float(input1_ptr[x]) * scale1 + float(input2_ptr[x]) * scale2 + offset;
-#ifdef __aarch64__
-                output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::lround(result));
-#else  // __aarch64__
-                output_ptr[x] = utility::clamp<int, int8_t>(support::cpp11::trunc(result));
-#endif  // __aarch64__
-            }
-        },
-        input1, input2, output);
-    }
+    add_sub_qasymm8_signed_neon(src0, src1, dst, policy, window, true /*is_addition*/);
 }
 } // namespace cpu
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/cpu/kernels/sub/neon/list.h b/src/cpu/kernels/sub/neon/list.h
index ac1346001a..f7e1a040bd 100644
--- a/src/cpu/kernels/sub/neon/list.h
+++ b/src/cpu/kernels/sub/neon/list.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,6 +35,8 @@ 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_fixedpoint);
+DECLARE_SUB_KERNEL(sub_qasymm8_signed_neon_fixedpoint);
 DECLARE_SUB_KERNEL(sub_qasymm8_neon);
 DECLARE_SUB_KERNEL(sub_qasymm8_signed_neon);
 DECLARE_SUB_KERNEL(sub_qsymm16_neon);
@@ -81,7 +83,8 @@ void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const
         Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
         Iterator output(dst, win);
 
-        execute_window_loop(win, [&](const Coordinates &)
+        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());
@@ -127,7 +130,8 @@ void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const
         Iterator input2(src1, input2_win);
         Iterator output(dst, win);
 
-        execute_window_loop(win, [&](const Coordinates &)
+        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());
diff --git a/src/cpu/kernels/sub/neon/qasymm8.cpp b/src/cpu/kernels/sub/neon/qasymm8.cpp
index 8f4cd8bdbb..ea6e5826dd 100644
--- a/src/cpu/kernels/sub/neon/qasymm8.cpp
+++ b/src/cpu/kernels/sub/neon/qasymm8.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,209 +21,22 @@
  * 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"
+#include "src/cpu/kernels/add/generic/neon/impl.h"
 
 namespace arm_compute
 {
 namespace cpu
 {
-void sub_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+void sub_qasymm8_neon_fixedpoint(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;
+    add_sub_q8_neon_fixedpoint<uint8_t>(src0, src1, dst, policy, window, false /*is_addition*/);
+}
 
-                *(output_ptr + x) = quantize_qasymm8((afs - bfs), dst->info()->quantization_info());
-            }
-        },
-        input1, input2, output);
-    }
+void sub_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    add_sub_qasymm8_neon(src0, src1, dst, policy, window, false /*is_addition*/);
 }
 
 } // namespace cpu
diff --git a/src/cpu/kernels/sub/neon/qasymm8_signed.cpp b/src/cpu/kernels/sub/neon/qasymm8_signed.cpp
index 2c9e411743..a86c7f22f6 100644
--- a/src/cpu/kernels/sub/neon/qasymm8_signed.cpp
+++ b/src/cpu/kernels/sub/neon/qasymm8_signed.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,209 +21,24 @@
  * 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"
+#include "src/cpu/kernels/add/generic/neon/impl.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)
+void sub_qasymm8_signed_neon_fixedpoint(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;
+    add_sub_q8_neon_fixedpoint<int8_t>(src0, src1, dst, policy, window, false /*is_addition*/);
+}
 
-                *(output_ptr + x) = quantize_qasymm8_signed((afs - bfs), dst->info()->quantization_info());
-            }
-        },
-        input1, input2, output);
-    }
+void sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    add_sub_qasymm8_signed_neon(src0, src1, dst, policy, window, false /*is_addition*/);
 }
+
 } // namespace cpu
 } // namespace arm_compute
\ No newline at end of file