COMPMID-3176: Remove padding from NEArithmeticSubtractionKernel

COMPMID-3487: Refactor NEArithmeticSubtractionKernel

Refactored code in order to remove paddings. This resulted in
a big increase in libary size so after some rework the total size
dropped by 4Kb.

Change-Id: I4e3014c2ae49c29c6090b195ea16620afcf6c09f
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3206
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 588 insertions, 430 deletions

diff --git a/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp b/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp
index 9b7b235c9f..8bfb37ea18 100644
--- a/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp
@@ -26,6 +26,7 @@
 #include "arm_compute/core/CPP/Validate.h"
 #include "arm_compute/core/NEON/NEAsymm.h"
 #include "arm_compute/core/NEON/NESymm.h"
+#include "arm_compute/core/NEON/wrapper/wrapper.h"
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Validate.h"
 
@@ -33,437 +34,628 @@ namespace arm_compute
 {
 namespace
 {
-constexpr unsigned int num_elems_processed_per_iteration = 16;
-
-void sub_wrap_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+template <typename T>
+inline typename std::enable_if<std::is_same<T, int8_t>::value, int8_t>::type
+quantize(float val, const QuantizationInfo &info)
 {
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
-
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        const uint8x16_t ta1 = vld1q_u8(input1.ptr());
-        const uint8x16_t ta2 = vld1q_u8(input2.ptr());
+    return quantize_qasymm8_signed(val, info);
+}
 
-        vst1q_u8(output.ptr(), vsubq_u8(ta1, ta2));
-    },
-    input1, input2, output);
+template <typename T>
+inline typename std::enable_if<std::is_same<T, uint8_t>::value, uint8_t>::type
+quantize(float val, const QuantizationInfo &info)
+{
+    return quantize_qasymm8(val, info);
 }
 
-void sub_saturate_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+template <typename T>
+void sub_same(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, bool is_sat)
 {
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
+    /** NEON vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
 
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        const uint8x16_t ta1 = vld1q_u8(input1.ptr());
-        const uint8x16_t ta2 = vld1q_u8(input2.ptr());
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
 
-        vst1q_u8(output.ptr(), vqsubq_u8(ta1, ta2));
-    },
-    input1, input2, output);
-}
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    constexpr int window_step_x         = 16 / sizeof(T);
+    const auto    window_start_x        = static_cast<int>(window.x().start());
+    const auto    window_end_x          = static_cast<int>(window.x().end());
+    const bool    is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
 
-void sub_saturate_QAYSMM8_QAYSMM8_QAYSMM8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
-{
     Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
     Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
     Iterator output(out, window);
 
-    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
-    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
-    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
-
-    execute_window_loop(window, [&](const Coordinates &)
+    if(is_broadcast_across_x)
     {
-        const float32x4x4_t ta1 = vdequantize(vld1q_u8(reinterpret_cast<const qasymm8_t *>(input1.ptr())), iq1_info);
-        const float32x4x4_t ta2 = vdequantize(vld1q_u8(reinterpret_cast<const qasymm8_t *>(input2.ptr())), iq2_info);
+        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 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
 
-        const float32x4x4_t ta3 =
-        {
-            {
-                vsubq_f32(ta1.val[0], ta2.val[0]),
-                vsubq_f32(ta1.val[1], ta2.val[1]),
-                vsubq_f32(ta1.val[2], ta2.val[2]),
-                vsubq_f32(ta1.val[3], ta2.val[3]),
-            }
-        };
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-        const uint8x16_t result = vquantize(ta3, oq_info);
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
 
-        vst1q_u8(reinterpret_cast<qasymm8_t *>(output.ptr()), result);
-    },
-    input1, input2, output);
-}
+        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());
 
-void sub_saturate_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
-{
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
+            const T    broadcast_value     = *reinterpret_cast<const T *>(broadcast_input.ptr());
+            const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
 
-    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
-    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
-    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
+                auto       res             = is_sat ? wrapper::vqsub(broadcast_value_vec, non_broadcast_v) : wrapper::vsub(broadcast_value_vec, non_broadcast_v);
+                if(is_broadcast_input_2)
+                {
+                    res = wrapper::vmul(res, wrapper::vdup_n(static_cast<T>(-1), ExactTagType{}));
+                }
+                wrapper::vstore(output_ptr + x, res);
+            }
 
-    execute_window_loop(window, [&](const Coordinates &)
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
+                auto       res             = is_sat ? wrapper::sub_sat(broadcast_value, non_broadcast_v) : broadcast_value - non_broadcast_v;
+                if(is_broadcast_input_2)
+                {
+                    res = static_cast<T>(-1) * res;
+                }
+
+                *(output_ptr + x) = res;
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
     {
-        const float32x4x4_t ta1 = vdequantize(vld1q_s8(reinterpret_cast<const qasymm8_signed_t *>(input1.ptr())), iq1_info);
-        const float32x4x4_t ta2 = vdequantize(vld1q_s8(reinterpret_cast<const qasymm8_signed_t *>(input2.ptr())), iq2_info);
+        // 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));
 
-        const float32x4x4_t ta3 =
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
         {
+            const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<T *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
             {
-                vsubq_f32(ta1.val[0], ta2.val[0]),
-                vsubq_f32(ta1.val[1], ta2.val[1]),
-                vsubq_f32(ta1.val[2], ta2.val[2]),
-                vsubq_f32(ta1.val[3], ta2.val[3]),
+                const auto val1 = wrapper::vloadq(input1_ptr + x);
+                const auto val2 = wrapper::vloadq(input2_ptr + x);
+                const auto res  = is_sat ? wrapper::vqsub(val1, val2) : wrapper::vsub(val1, val2);
+                wrapper::vstore(output_ptr + x, res);
             }
-        };
 
-        const int8x16_t result = vquantize_signed(ta3, oq_info);
-
-        vst1q_s8(reinterpret_cast<qasymm8_signed_t *>(output.ptr()), result);
-    },
-    input1, input2, output);
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto val1   = *(input1_ptr + x);
+                const auto val2   = *(input2_ptr + x);
+                *(output_ptr + x) = is_sat ? wrapper::sub_sat(val1, val2) : val1 - val2;
+            }
+        },
+        input1, input2, output);
+    }
 }
 
-void sub_saturate_QSYMM16_QSYMM16_QSYMM16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+template <typename T>
+void sub_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, bool is_sat)
 {
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
+    ARM_COMPUTE_UNUSED(is_sat);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->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 = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
 
     const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
     const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
     const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
 
-    execute_window_loop(window, [&](const Coordinates &)
+    const float32x4_t vscale1    = vdupq_n_f32(iq1_info.scale);
+    const float32x4_t vscale2    = vdupq_n_f32(iq2_info.scale);
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
+    const int32x4_t   voffset1   = vdupq_n_s32(iq1_info.offset);
+    const int32x4_t   voffset2   = vdupq_n_s32(iq2_info.offset);
+    const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
+
+    if(is_broadcast_across_x)
     {
-        const int16x8x2_t in1_s16 =
+        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 ? in2 : in1;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
         {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<T *>(output.ptr());
+
+            const auto broadcast_value     = *reinterpret_cast<const T *>(broadcast_input.ptr());
+            const auto broadcast_value_vec = wrapper::vdup_n(static_cast<T>(broadcast_value), wrapper::traits::vector_128_tag{});
+
+            const float32x4x4_t bf =
             {
-                vld1q_s16(reinterpret_cast<const qsymm16_t *>(input1.ptr())),
-                vld1q_s16(reinterpret_cast<const qsymm16_t *>(input1.ptr()) + 8),
-            }
-        };
-        const int16x8x2_t in2_s16 =
-        {
+                {
+                    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),
+                }
+            };
+            const float bfs = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
             {
-                vld1q_s16(reinterpret_cast<const qsymm16_t *>(input2.ptr())),
-                vld1q_s16(reinterpret_cast<const qsymm16_t *>(input2.ptr()) + 8),
+                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 = wrapper::vqmov<T>(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
+                const auto pb = wrapper::vqmov<T>(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
+                wrapper::vstore(output_ptr + x, wrapper::vcombine(pa, pb));
             }
-        };
-        const float32x4x4_t ta1 = vdequantize(in1_s16, iq1_info);
-        const float32x4x4_t ta2 = vdequantize(in2_s16, iq2_info);
 
-        const float32x4x4_t ta3 =
-        {
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
             {
-                vsubq_f32(ta1.val[0], ta2.val[0]),
-                vsubq_f32(ta1.val[1], ta2.val[1]),
-                vsubq_f32(ta1.val[2], ta2.val[2]),
-                vsubq_f32(ta1.val[3], ta2.val[3]),
+                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
+                *(output_ptr + x) = quantize<T>((afs - bfs), out->info()->quantization_info());
             }
-        };
-
-        const int16x8x2_t result = vquantize_qsymm16(ta3, oq_info);
-
-        vst1q_s16(reinterpret_cast<qsymm16_t *>(output.ptr()), result.val[0]);
-        vst1q_s16(reinterpret_cast<qsymm16_t *>(output.ptr()) + 8, result.val[1]);
-    },
-    input1, input2, output);
-}
+        },
+        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));
 
-void sub_wrap_S16_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
-{
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
 
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        const int16x8x2_t ta1 =
-        {
-            {
-                vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr())),
-                vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr()) + 8),
-            }
-        };
-        const int16x8x2_t ta2 =
+        execute_window_loop(win, [&](const Coordinates &)
         {
+            const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<T *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
             {
-                vld1q_s16(reinterpret_cast<const int16_t *>(input2.ptr())),
-                vld1q_s16(reinterpret_cast<const int16_t *>(input2.ptr()) + 8),
+                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 = wrapper::vqmov<T>(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
+                const auto pb = wrapper::vqmov<T>(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
+                wrapper::vstore(output_ptr + x, wrapper::vcombine(pa, pb));
             }
-        };
 
-        const int16x8x2_t ta3 =
-        {
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
             {
-                vsubq_s16(ta1.val[0], ta2.val[0]),
-                vsubq_s16(ta1.val[1], ta2.val[1])
-            }
-        };
+                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;
 
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), ta3.val[0]);
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, ta3.val[1]);
-    },
-    input1, input2, output);
+                *(output_ptr + x) = quantize<T>((afs - bfs), out->info()->quantization_info());
+            }
+        },
+        input1, input2, output);
+    }
 }
 
-void sub_saturate_S16_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+void sub_QSYMM16_QSYMM16_QSYMM16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, bool is_sat)
 {
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
+    ARM_COMPUTE_UNUSED(is_sat);
+
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x         = 8;
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
+
+    const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
+    const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
+    const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
+
+    const float32x4_t vscale1    = vdupq_n_f32(iq1_info.scale);
+    const float32x4_t vscale2    = vdupq_n_f32(iq2_info.scale);
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
 
-    execute_window_loop(window, [&](const Coordinates &)
+    if(is_broadcast_across_x)
     {
-        const int16x8x2_t ta1 =
+        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 ? in2 : in1;
+        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
         {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int16_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int16_t *>(output.ptr());
+
+            const int16_t   broadcast_value     = *reinterpret_cast<const int16_t *>(broadcast_input.ptr());
+            const int16x8_t broadcast_value_vec = vdupq_n_s16(broadcast_value);
+
+            const float32x4x2_t bf =
             {
-                vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr())),
-                vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr()) + 8),
-            }
-        };
-        const int16x8x2_t ta2 =
-        {
+                {
+                    vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(broadcast_value_vec))), vscale2),
+                    vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(broadcast_value_vec))), vscale2),
+                }
+            };
+            const float bfs = static_cast<int32_t>(broadcast_value) * broadcast_qinfo.scale;
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
             {
-                vld1q_s16(reinterpret_cast<const int16_t *>(input2.ptr())),
-                vld1q_s16(reinterpret_cast<const int16_t *>(input2.ptr()) + 8),
+                const int16x8_t     a = vld1q_s16(non_broadcast_input_ptr + x);
+                const float32x4x2_t af =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1),
+                    }
+                };
+
+                const int32x4x4_t rf =
+                {
+                    {
+#ifdef __aarch64__
+                        vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+#else  //__aarch64__
+                        vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtq_s32_f32(vmulq_f32(is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+#endif //__aarch64__
+                    }
+                };
+
+                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1]));
+                vst1q_s16(output_ptr + x, pa);
             }
-        };
 
-        const int16x8x2_t ta3 =
-        {
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
             {
-                vqsubq_s16(ta1.val[0], ta2.val[0]),
-                vqsubq_s16(ta1.val[1], ta2.val[1])
+                const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x)) * non_broadcast_qinfo.scale;
+                *(output_ptr + x) = quantize_qsymm16(is_broadcast_input_2 ? (bfs - afs) : (afs - bfs), oq_info);
             }
-        };
-
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), ta3.val[0]);
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, ta3.val[1]);
-    },
-    input1, input2, output);
-}
+        },
+        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));
 
-void sub_F16_F16_F16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
-{
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
 
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        const float16x8x2_t a =
-        {
-            {
-                vld1q_f16(reinterpret_cast<const float16_t *>(input1.ptr())),
-                vld1q_f16(reinterpret_cast<const float16_t *>(input1.ptr()) + 8),
-            }
-        };
-        const float16x8x2_t b =
+        execute_window_loop(win, [&](const Coordinates &)
         {
+            const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int16_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
             {
-                vld1q_f16(reinterpret_cast<const float16_t *>(input2.ptr())),
-                vld1q_f16(reinterpret_cast<const float16_t *>(input2.ptr()) + 8),
+                const int16x8_t a = vld1q_s16(input1_ptr + x);
+                const int16x8_t b = vld1q_s16(input2_ptr + x);
+
+                const float32x4x2_t af =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(a))), vscale1),
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(a))), vscale1),
+                    }
+                };
+
+                const float32x4x2_t bf =
+                {
+                    {
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(b))), vscale2),
+                        vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(b))), vscale2),
+                    }
+                };
+
+                const int32x4x2_t rf =
+                {
+                    {
+#ifdef __aarch64__
+                        vcvtnq_s32_f32(vmulq_f32(vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmulq_f32(vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+#else  //__aarch64__
+                        vcvtq_s32_f32(vmulq_f32(vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtq_s32_f32(vmulq_f32(vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
+#endif //__aarch64__
+                    }
+                };
+
+                const int16x8_t pa = vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1]));
+                vst1q_s16(output_ptr + x, pa);
             }
-        };
-        const float16x8x2_t res =
-        {
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
             {
-                vsubq_f16(a.val[0], b.val[0]),
-                vsubq_f16(a.val[1], b.val[1]),
+                const float afs   = static_cast<int32_t>((*(input1_ptr + x))) * iq1_info.scale;
+                const float bfs   = static_cast<int32_t>((*(input2_ptr + x))) * iq2_info.scale;
+                *(output_ptr + x) = quantize_qsymm16((afs - bfs), out->info()->quantization_info());
             }
-        };
-
-        vst1q_f16(reinterpret_cast<float16_t *>(output.ptr()), res.val[0]);
-        vst1q_f16(reinterpret_cast<float16_t *>(output.ptr()) + 8, res.val[1]);
-    },
-    input1, input2, output);
-#else  /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-    ARM_COMPUTE_UNUSED(in1);
-    ARM_COMPUTE_UNUSED(in2);
-    ARM_COMPUTE_UNUSED(out);
-    ARM_COMPUTE_UNUSED(window);
-    ARM_COMPUTE_ERROR("Not supported, recompile the library with arch=arm64-v8.2-a");
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+        },
+        input1, input2, output);
+    }
 }
 
-void sub_F32_F32_F32(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+void sub_S16_U8_S16_impl(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, bool is_sat, bool is_swapped)
 {
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
 
-    execute_window_loop(window, [&](const Coordinates &)
+    // Clear X Dimension on execution window as we handle manually
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input1(in1, input1_win);
+    Iterator input2(in2, input2_win);
+    Iterator output(out, win);
+
+    const int  window_step_x  = 8;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    execute_window_loop(win, [&](const Coordinates &)
     {
-        const float32x4x4_t ta1 =
+        const auto input1_ptr = reinterpret_cast<const int16_t *>(input1.ptr());
+        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+        if(!is_sat)
         {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
             {
-                vld1q_f32(reinterpret_cast<const float *>(input1.ptr())),
-                vld1q_f32(reinterpret_cast<const float *>(input1.ptr()) + 4),
-                vld1q_f32(reinterpret_cast<const float *>(input1.ptr()) + 8),
-                vld1q_f32(reinterpret_cast<const float *>(input1.ptr()) + 12),
+                const auto vin1 = wrapper::vloadq(input1_ptr + x);
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                const auto res  = is_swapped ? wrapper::vsub(vin2, vin1) : wrapper::vsub(vin1, vin2);
+                wrapper::vstore(output_ptr + x, res);
             }
-        };
-        const float32x4x4_t ta2 =
-        {
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
             {
-                vld1q_f32(reinterpret_cast<const float *>(input2.ptr())),
-                vld1q_f32(reinterpret_cast<const float *>(input2.ptr()) + 4),
-                vld1q_f32(reinterpret_cast<const float *>(input2.ptr()) + 8),
-                vld1q_f32(reinterpret_cast<const float *>(input2.ptr()) + 12),
+                const auto res    = is_swapped ? static_cast<int16_t>(*(input2_ptr + x)) - *(input1_ptr + x) : *(input1_ptr + x) - static_cast<int16_t>(*(input2_ptr + x));
+                *(output_ptr + x) = res;
             }
-        };
-
-        const float32x4x4_t ta3 =
+        }
+        else
         {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
             {
-                vsubq_f32(ta1.val[0], ta2.val[0]),
-                vsubq_f32(ta1.val[1], ta2.val[1]),
-                vsubq_f32(ta1.val[2], ta2.val[2]),
-                vsubq_f32(ta1.val[3], ta2.val[3]),
+                const auto vin1 = wrapper::vloadq(input1_ptr + x);
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                const auto res  = is_swapped ? wrapper::vqsub(vin2, vin1) : wrapper::vqsub(vin1, vin2);
+                wrapper::vstore(output_ptr + x, res);
             }
-        };
-
-        vst1q_f32(reinterpret_cast<float *>(output.ptr()), ta3.val[0]);
-        vst1q_f32(reinterpret_cast<float *>(output.ptr()) + 4, ta3.val[1]);
-        vst1q_f32(reinterpret_cast<float *>(output.ptr()) + 8, ta3.val[2]);
-        vst1q_f32(reinterpret_cast<float *>(output.ptr()) + 12, ta3.val[3]);
-    },
-    input1, input2, output);
-}
-void sub_wrap_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
-{
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
 
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        const uint8x16_t bv_0 = vld1q_u8(input2.ptr());
-        int16x8_t        a1_0 = vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr()));
-        int16x8_t        a2_0 = vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr()) + 8);
-
-        a1_0 = vsubq_s16(a1_0, vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bv_0))));
-        a2_0 = vsubq_s16(a2_0, vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bv_0))));
-
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), a1_0);
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, a2_0);
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                const auto res    = is_swapped ? wrapper::sub_sat(static_cast<int16_t>(*(input2_ptr + x)), *(input1_ptr + x)) : wrapper::sub_sat(*(input1_ptr + x), static_cast<int16_t>(*(input2_ptr + x)));
+                *(output_ptr + x) = res;
+            }
+        }
     },
     input1, input2, output);
 }
 
-void sub_saturate_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+void sub_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, bool is_sat)
 {
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
-
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        const uint8x16_t bv_0 = vld1q_u8(input2.ptr());
-        int16x8_t        a1_0 = vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr()));
-        int16x8_t        a2_0 = vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr()) + 8);
-
-        a1_0 = vqsubq_s16(a1_0, vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bv_0))));
-        a2_0 = vqsubq_s16(a2_0, vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bv_0))));
-
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), a1_0);
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, a2_0);
-    },
-    input1, input2, output);
+    sub_S16_U8_S16_impl(in1, in2, out, window, is_sat, false);
 }
 
-void sub_wrap_U8_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+void sub_U8_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, bool is_sat)
 {
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
-
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        const uint8x16_t bv_0 = vld1q_u8(input1.ptr());
-        int16x8_t        a1_0 = vld1q_s16(reinterpret_cast<const int16_t *>(input2.ptr()));
-        int16x8_t        a2_0 = vld1q_s16(reinterpret_cast<const int16_t *>(input2.ptr()) + 8);
-
-        a1_0 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bv_0))), a1_0);
-        a2_0 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bv_0))), a2_0);
-
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), a1_0);
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, a2_0);
-    },
-    input1, input2, output);
+    // Swap arguments
+    sub_S16_U8_S16_impl(in2, in1, out, window, is_sat, true);
 }
 
-void sub_saturate_U8_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+void sub_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, bool is_sat)
 {
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
+    // Create input windows
+    Window win        = window;
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
 
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        const uint8x16_t bv_0 = vld1q_u8(input1.ptr());
-        int16x8_t        a1_0 = vld1q_s16(reinterpret_cast<const int16_t *>(input2.ptr()));
-        int16x8_t        a2_0 = vld1q_s16(reinterpret_cast<const int16_t *>(input2.ptr()) + 8);
+    // Clear X Dimension on execution window as we handle manually
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-        a1_0 = vqsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bv_0))), a1_0);
-        a2_0 = vqsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bv_0))), a2_0);
+    Iterator input1(in1, input1_win);
+    Iterator input2(in2, input2_win);
+    Iterator output(out, win);
 
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), a1_0);
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, a2_0);
-    },
-    input1, input2, output);
-}
+    const int  window_step_x  = 8;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
 
-void sub_wrap_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
-{
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
-
-    execute_window_loop(window, [&](const Coordinates &)
+    execute_window_loop(win, [&](const Coordinates &)
     {
-        const uint8x16_t av_0 = vld1q_u8(input1.ptr());
-        const uint8x16_t bv_0 = vld1q_u8(input2.ptr());
+        const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());
+        const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.ptr());
+        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
 
-        const int16x8_t a1_0 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(av_0))),
-                                         vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bv_0))));
-        const int16x8_t a2_0 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(av_0))),
-                                         vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bv_0))));
-
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), a1_0);
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, a2_0);
-    },
-    input1, input2, output);
-}
-
-void sub_saturate_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
-{
-    Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
-    Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
-    Iterator output(out, window);
-
-    execute_window_loop(window, [&](const Coordinates &)
-    {
-        const uint8x16_t av_0 = vld1q_u8(input1.ptr());
-        const uint8x16_t bv_0 = vld1q_u8(input2.ptr());
+        if(!is_sat)
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                wrapper::vstore(output_ptr + x, wrapper::vsub(vin1, vin2));
+            }
 
-        const int16x8_t a1_0 = vqsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(av_0))),
-                                          vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bv_0))));
-        const int16x8_t a2_0 = vqsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(av_0))),
-                                          vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bv_0))));
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + x) = static_cast<int16_t>(*(input1_ptr + x)) - static_cast<int16_t>(*(input2_ptr + x));
+            }
+        }
+        else
+        {
+            // Compute S elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x)));
+                const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x)));
+                wrapper::vstore(output_ptr + x, wrapper::vqsub(vin1, vin2));
+            }
 
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), a1_0);
-        vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, a2_0);
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + x) = wrapper::sub_sat(static_cast<int16_t>(*(input1_ptr + x)),
+                                                     static_cast<int16_t>(*(input2_ptr + x)));
+            }
+        }
     },
     input1, input2, output);
 }
@@ -519,64 +711,10 @@ inline Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &i
     }
     return Status{};
 }
-
-inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output)
-{
-    const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(input1, input2);
-    const TensorShape &out_shape    = broadcast_pair.first;
-    const ValidRegion &valid_region = broadcast_pair.second;
-
-    // Auto initialize output if not initialized
-    {
-        set_shape_if_empty(output, out_shape);
-
-        if(input1.data_type() == DataType::S16 || input2.data_type() == DataType::S16)
-        {
-            set_format_if_unknown(output, Format::S16);
-        }
-        else if(input1.data_type() == DataType::F16 || input2.data_type() == DataType::F16)
-        {
-            set_format_if_unknown(output, Format::F16);
-        }
-        else if(input1.data_type() == DataType::F32 || input2.data_type() == DataType::F32)
-        {
-            set_format_if_unknown(output, Format::F32);
-        }
-        else if(input1.data_type() == DataType::QASYMM8 || input2.data_type() == DataType::QASYMM8)
-        {
-            set_data_type_if_unknown(output, DataType::QASYMM8);
-        }
-        else if(input1.data_type() == DataType::QASYMM8_SIGNED || input2.data_type() == DataType::QASYMM8_SIGNED)
-        {
-            set_data_type_if_unknown(output, DataType::QASYMM8_SIGNED);
-        }
-        else if(input1.data_type() == DataType::QSYMM16 || input2.data_type() == DataType::QSYMM16)
-        {
-            set_data_type_if_unknown(output, DataType::QSYMM16);
-        }
-    }
-
-    Window win        = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration));
-    Window win_input1 = win.broadcast_if_dimension_le_one(input1);
-    Window win_input2 = win.broadcast_if_dimension_le_one(input2);
-
-    AccessWindowHorizontal input1_access(&input1, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal input2_access(&input2, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_access(&output, 0, num_elems_processed_per_iteration);
-
-    bool window_changed = update_window_and_padding(win_input1, input1_access)
-                          || update_window_and_padding(win_input2, input2_access)
-                          || update_window_and_padding(win, output_access);
-
-    output_access.set_valid_region(win, valid_region);
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
-}
 } // namespace
 
 NEArithmeticSubtractionKernel::NEArithmeticSubtractionKernel()
-    : _func(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr)
+    : _func(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _policy(ConvertPolicy::WRAP)
 {
 }
 
@@ -585,57 +723,84 @@ void NEArithmeticSubtractionKernel::configure(const ITensor *input1, const ITens
     ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info(), policy));
 
-    // Configure kernel window
-    auto win_config = validate_and_configure_window(*input1->info(), *input2->info(), *output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-
-    static std::map<std::string, NEArithmeticSubtractionKernel::SubFunction *> map_function =
-    {
-        { "sub_wrap_U8_U8_U8", &sub_wrap_U8_U8_U8 },
-        { "sub_wrap_U8_U8_S16", &sub_wrap_U8_U8_S16 },
-        { "sub_saturate_U8_U8_U8", &sub_saturate_U8_U8_U8 },
-        { "sub_saturate_U8_U8_S16", &sub_saturate_U8_U8_S16 },
-        { "sub_saturate_QASYMM8_QASYMM8_QASYMM8", &sub_saturate_QAYSMM8_QAYSMM8_QAYSMM8 },
-        { "sub_saturate_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &sub_saturate_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED },
-        { "sub_saturate_QSYMM16_QSYMM16_QSYMM16", &sub_saturate_QSYMM16_QSYMM16_QSYMM16 },
-        { "sub_wrap_U8_S16_S16", &sub_wrap_U8_S16_S16 },
-        { "sub_wrap_S16_U8_S16", &sub_wrap_S16_U8_S16 },
-        { "sub_saturate_U8_S16_S16", &sub_saturate_U8_S16_S16 },
-        { "sub_saturate_S16_U8_S16", &sub_saturate_S16_U8_S16 },
-        { "sub_wrap_S16_S16_S16", &sub_wrap_S16_S16_S16 },
-        { "sub_saturate_S16_S16_S16", &sub_saturate_S16_S16_S16 },
-        { "sub_wrap_F32_F32_F32", &sub_F32_F32_F32 },
-        { "sub_saturate_F32_F32_F32", &sub_F32_F32_F32 },
-        { "sub_wrap_F16_F16_F16", &sub_F16_F16_F16 },
-        { "sub_saturate_F16_F16_F16", &sub_F16_F16_F16 },
-    };
-
     _input1 = input1;
     _input2 = input2;
     _output = output;
+    _policy = policy;
 
-    std::string function_to_call("sub_");
-    function_to_call += policy == ConvertPolicy::WRAP ? "wrap_" : "saturate_";
-    function_to_call += string_from_data_type(input1->info()->data_type()) + "_";
-    function_to_call += string_from_data_type(input2->info()->data_type()) + "_";
-    function_to_call += string_from_data_type(output->info()->data_type());
+    const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1->info(), *input2->info());
+    const TensorShape &out_shape    = broadcast_pair.first;
+    const ValidRegion &valid_region = broadcast_pair.second;
 
-    auto it = map_function.find(function_to_call);
+    // Auto initialize output if not initialized
+    set_shape_if_empty(*output->info(), out_shape);
 
-    if(it != map_function.end())
+    switch(input1->info()->data_type())
     {
-        _func = it->second;
+        case DataType::U8:
+            if(input2->info()->data_type() == DataType::U8 && output->info()->data_type() == DataType::U8)
+            {
+                _func = &sub_same<uint8_t>;
+            }
+            else if(input2->info()->data_type() == DataType::U8 && output->info()->data_type() == DataType::S16)
+            {
+                _func = &sub_U8_U8_S16;
+            }
+            else
+            {
+                _func = &sub_U8_S16_S16;
+            }
+            break;
+        case DataType::QASYMM8:
+            _func = &sub_quantized<uint8_t>;
+            set_data_type_if_unknown(*output->info(), DataType::QASYMM8);
+            break;
+        case DataType::QASYMM8_SIGNED:
+            _func = &sub_quantized<int8_t>;
+            set_data_type_if_unknown(*output->info(), DataType::QASYMM8_SIGNED);
+            break;
+        case DataType::S16:
+            if(input2->info()->data_type() == DataType::U8)
+            {
+                _func = &sub_S16_U8_S16;
+            }
+            else
+            {
+                _func = &sub_same<int16_t>;
+            }
+            set_format_if_unknown(*output->info(), Format::S16);
+            break;
+        case DataType::QSYMM16:
+            _func = &sub_QSYMM16_QSYMM16_QSYMM16;
+            set_data_type_if_unknown(*output->info(), DataType::QSYMM16);
+            break;
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        case DataType::F16:
+            _func = &sub_same<float16_t>;
+            set_format_if_unknown(*output->info(), Format::F16);
+            break;
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+        case DataType::F32:
+            _func = &sub_same<float>;
+            set_format_if_unknown(*output->info(), Format::F32);
+            break;
+        default:
+            _func = nullptr;
     }
 
-    INEKernel::configure(win_config.second);
+    // NEArithmeticSubtractionKernel doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(output->info()->num_dimensions());
+    output->info()->set_valid_region(valid_region);
+    Window win = calculate_max_window(valid_region, Steps());
+
+    INEKernel::configure(win);
 }
 
 Status NEArithmeticSubtractionKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
-
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output, policy));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(*input1->clone(), *input2->clone(), *output->clone()).first);
 
     return Status{};
 }
@@ -647,13 +812,6 @@ void NEArithmeticSubtractionKernel::run(const Window &window, const ThreadInfo &
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
     ARM_COMPUTE_ERROR_ON(_func == nullptr);
 
-    (*_func)(_input1, _input2, _output, window);
-}
-
-BorderSize NEArithmeticSubtractionKernel::border_size() const
-{
-    const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0));
-    const unsigned int border        = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize);
-    return BorderSize{ 0, border, 0, 0 };
+    (*_func)(_input1, _input2, _output, window, (_policy == ConvertPolicy::SATURATE));
 }
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute