COMPMID-675 - Reworked NEGEMMLowp interface/function

The new interface makes NEGEMMLowp able to work with ASYMM8 data types.

Implemented 2 new functions:
- NEGEMMLowpMatrixMultiplyCore
- NEGEMMLowpOutputStage

These functions should make the integration in android NN doable

For more information about GEMMLowp:
https://github.com/google/gemmlowp/blob/master/doc/low-precision.md

Change-Id: Ie2c775f45234f68ca53dba644b3a912b997fd890
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/95504
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

6 files changed, 437 insertions, 465 deletions

diff --git a/src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.cpp b/src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.cpp
index ae5d456141..a29b661a00 100644
--- a/src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.cpp
@@ -132,7 +132,8 @@ NEGEMMInterleave4x4Kernel::NEGEMMInterleave4x4Kernel()
 
 void NEGEMMInterleave4x4Kernel::configure(const ITensor *input, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::U8, DataType::S8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16,
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::U8, DataType::S8, DataType::U16, DataType::S16, DataType::U32, DataType::S32,
+                                                  DataType::F16,
                                                   DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
diff --git a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
index 4b9c9f3e64..1352f34e3c 100644
--- a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
@@ -52,7 +52,7 @@ NEGEMMLowpMatrixMultiplyKernel::NEGEMMLowpMatrixMultiplyKernel()
 
 void NEGEMMLowpMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::S8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
 
@@ -127,115 +127,115 @@ void NEGEMMLowpMatrixMultiplyKernel::run(const Window &window, const ThreadInfo
     // All the values needed for computing a single 4x4 block will be read from consecutive memory positions
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        auto *mtx_a0 = reinterpret_cast<const int8_t *>(ina.ptr());
-        auto *mtx_b0 = reinterpret_cast<const int8_t *>(inb.ptr());
+        const uint8_t *mtx_a0 = ina.ptr();
+        const uint8_t *mtx_b0 = inb.ptr();
 
         // Note: Since the input are all positives, we can use uint32_t
         // Accumulators for the block 0
-        int32x4x4_t c0 =
+        uint32x4x4_t c0 =
         {
             {
-                vdupq_n_s32(0),
-                vdupq_n_s32(0),
-                vdupq_n_s32(0),
-                vdupq_n_s32(0)
+                vdupq_n_u32(0),
+                vdupq_n_u32(0),
+                vdupq_n_u32(0),
+                vdupq_n_u32(0)
             }
         };
 
         // Accumulators for the block 1
-        int32x4x4_t c1 =
+        uint32x4x4_t c1 =
         {
             {
-                vdupq_n_s32(0),
-                vdupq_n_s32(0),
-                vdupq_n_s32(0),
-                vdupq_n_s32(0)
+                vdupq_n_u32(0),
+                vdupq_n_u32(0),
+                vdupq_n_u32(0),
+                vdupq_n_u32(0)
             }
         };
 
         // Accumulators for the block 2
-        int32x4x4_t c2 =
+        uint32x4x4_t c2 =
         {
             {
-                vdupq_n_s32(0),
-                vdupq_n_s32(0),
-                vdupq_n_s32(0),
-                vdupq_n_s32(0)
+                vdupq_n_u32(0),
+                vdupq_n_u32(0),
+                vdupq_n_u32(0),
+                vdupq_n_u32(0)
             }
         };
 
         // Accumulators for the block 3
-        int32x4x4_t c3 =
+        uint32x4x4_t c3 =
         {
             {
-                vdupq_n_s32(0),
-                vdupq_n_s32(0),
-                vdupq_n_s32(0),
-                vdupq_n_s32(0)
+                vdupq_n_u32(0),
+                vdupq_n_u32(0),
+                vdupq_n_u32(0),
+                vdupq_n_u32(0)
             }
         };
 
         for(int k = 0; k < width_b; k += 16, mtx_a0 += 4, mtx_b0 += 16)
         {
-            const int8x8_t  a00_s8 = vld1_s8(mtx_a0);
-            const int8x16_t b00_s8 = vld1q_s8(mtx_b0);
+            const uint8x8_t  a00_u8 = vld1_u8(mtx_a0);
+            const uint8x16_t b00_u8 = vld1q_u8(mtx_b0);
 
             // Convert a00_s8 to uint16_t and get the lower part
-            const int16x4_t a00_s16 = vget_low_s16(vmovl_s8(a00_s8));
+            const uint16x4_t a00_u16 = vget_low_u16(vmovl_u8(a00_u8));
 
-            // Convert b00_s8 to int16_t
-            const int16x4x4_t b00_s16 =
+            // Convert b00_s8 to uint16_t
+            const uint16x4x4_t b00_u16 =
             {
                 {
-                    vget_low_s16(vmovl_s8(vget_low_s8(b00_s8))),
-                    vget_high_s16(vmovl_s8(vget_low_s8(b00_s8))),
-                    vget_low_s16(vmovl_s8(vget_high_s8(b00_s8))),
-                    vget_high_s16(vmovl_s8(vget_high_s8(b00_s8)))
+                    vget_low_u16(vmovl_u8(vget_low_u8(b00_u8))),
+                    vget_high_u16(vmovl_u8(vget_low_u8(b00_u8))),
+                    vget_low_u16(vmovl_u8(vget_high_u8(b00_u8))),
+                    vget_high_u16(vmovl_u8(vget_high_u8(b00_u8)))
                 }
             };
 
             // 4x4 block 0
-            c0.val[0] = vmlal_lane_s16(c0.val[0], b00_s16.val[0], a00_s16, 0);
-            c0.val[1] = vmlal_lane_s16(c0.val[1], b00_s16.val[1], a00_s16, 0);
-            c0.val[2] = vmlal_lane_s16(c0.val[2], b00_s16.val[2], a00_s16, 0);
-            c0.val[3] = vmlal_lane_s16(c0.val[3], b00_s16.val[3], a00_s16, 0);
+            c0.val[0] = vmlal_lane_u16(c0.val[0], b00_u16.val[0], a00_u16, 0);
+            c0.val[1] = vmlal_lane_u16(c0.val[1], b00_u16.val[1], a00_u16, 0);
+            c0.val[2] = vmlal_lane_u16(c0.val[2], b00_u16.val[2], a00_u16, 0);
+            c0.val[3] = vmlal_lane_u16(c0.val[3], b00_u16.val[3], a00_u16, 0);
 
             // 4x4 block 1
-            c1.val[0] = vmlal_lane_s16(c1.val[0], b00_s16.val[0], a00_s16, 1);
-            c1.val[1] = vmlal_lane_s16(c1.val[1], b00_s16.val[1], a00_s16, 1);
-            c1.val[2] = vmlal_lane_s16(c1.val[2], b00_s16.val[2], a00_s16, 1);
-            c1.val[3] = vmlal_lane_s16(c1.val[3], b00_s16.val[3], a00_s16, 1);
+            c1.val[0] = vmlal_lane_u16(c1.val[0], b00_u16.val[0], a00_u16, 1);
+            c1.val[1] = vmlal_lane_u16(c1.val[1], b00_u16.val[1], a00_u16, 1);
+            c1.val[2] = vmlal_lane_u16(c1.val[2], b00_u16.val[2], a00_u16, 1);
+            c1.val[3] = vmlal_lane_u16(c1.val[3], b00_u16.val[3], a00_u16, 1);
 
             // 4x4 block 2
-            c2.val[0] = vmlal_lane_s16(c2.val[0], b00_s16.val[0], a00_s16, 2);
-            c2.val[1] = vmlal_lane_s16(c2.val[1], b00_s16.val[1], a00_s16, 2);
-            c2.val[2] = vmlal_lane_s16(c2.val[2], b00_s16.val[2], a00_s16, 2);
-            c2.val[3] = vmlal_lane_s16(c2.val[3], b00_s16.val[3], a00_s16, 2);
+            c2.val[0] = vmlal_lane_u16(c2.val[0], b00_u16.val[0], a00_u16, 2);
+            c2.val[1] = vmlal_lane_u16(c2.val[1], b00_u16.val[1], a00_u16, 2);
+            c2.val[2] = vmlal_lane_u16(c2.val[2], b00_u16.val[2], a00_u16, 2);
+            c2.val[3] = vmlal_lane_u16(c2.val[3], b00_u16.val[3], a00_u16, 2);
 
             // 4x4 block 3
-            c3.val[0] = vmlal_lane_s16(c3.val[0], b00_s16.val[0], a00_s16, 3);
-            c3.val[1] = vmlal_lane_s16(c3.val[1], b00_s16.val[1], a00_s16, 3);
-            c3.val[2] = vmlal_lane_s16(c3.val[2], b00_s16.val[2], a00_s16, 3);
-            c3.val[3] = vmlal_lane_s16(c3.val[3], b00_s16.val[3], a00_s16, 3);
+            c3.val[0] = vmlal_lane_u16(c3.val[0], b00_u16.val[0], a00_u16, 3);
+            c3.val[1] = vmlal_lane_u16(c3.val[1], b00_u16.val[1], a00_u16, 3);
+            c3.val[2] = vmlal_lane_u16(c3.val[2], b00_u16.val[2], a00_u16, 3);
+            c3.val[3] = vmlal_lane_u16(c3.val[3], b00_u16.val[3], a00_u16, 3);
         }
 
         auto mtx_out = reinterpret_cast<int32_t *>(out.ptr());
-        vst1q_s32(mtx_out + 0 * out_stride + 0, c0.val[0]);
-        vst1q_s32(mtx_out + 0 * out_stride + 4, c0.val[1]);
-        vst1q_s32(mtx_out + 0 * out_stride + 8, c0.val[2]);
-        vst1q_s32(mtx_out + 0 * out_stride + 12, c0.val[3]);
-        vst1q_s32(mtx_out + 1 * out_stride + 0, c1.val[0]);
-        vst1q_s32(mtx_out + 1 * out_stride + 4, c1.val[1]);
-        vst1q_s32(mtx_out + 1 * out_stride + 8, c1.val[2]);
-        vst1q_s32(mtx_out + 1 * out_stride + 12, c1.val[3]);
-        vst1q_s32(mtx_out + 2 * out_stride + 0, c2.val[0]);
-        vst1q_s32(mtx_out + 2 * out_stride + 4, c2.val[1]);
-        vst1q_s32(mtx_out + 2 * out_stride + 8, c2.val[2]);
-        vst1q_s32(mtx_out + 2 * out_stride + 12, c2.val[3]);
-        vst1q_s32(mtx_out + 3 * out_stride + 0, c3.val[0]);
-        vst1q_s32(mtx_out + 3 * out_stride + 4, c3.val[1]);
-        vst1q_s32(mtx_out + 3 * out_stride + 8, c3.val[2]);
-        vst1q_s32(mtx_out + 3 * out_stride + 12, c3.val[3]);
+        vst1q_s32(mtx_out + 0 * out_stride + 0, vreinterpretq_s32_u32(c0.val[0]));
+        vst1q_s32(mtx_out + 0 * out_stride + 4, vreinterpretq_s32_u32(c0.val[1]));
+        vst1q_s32(mtx_out + 0 * out_stride + 8, vreinterpretq_s32_u32(c0.val[2]));
+        vst1q_s32(mtx_out + 0 * out_stride + 12, vreinterpretq_s32_u32(c0.val[3]));
+        vst1q_s32(mtx_out + 1 * out_stride + 0, vreinterpretq_s32_u32(c1.val[0]));
+        vst1q_s32(mtx_out + 1 * out_stride + 4, vreinterpretq_s32_u32(c1.val[1]));
+        vst1q_s32(mtx_out + 1 * out_stride + 8, vreinterpretq_s32_u32(c1.val[2]));
+        vst1q_s32(mtx_out + 1 * out_stride + 12, vreinterpretq_s32_u32(c1.val[3]));
+        vst1q_s32(mtx_out + 2 * out_stride + 0, vreinterpretq_s32_u32(c2.val[0]));
+        vst1q_s32(mtx_out + 2 * out_stride + 4, vreinterpretq_s32_u32(c2.val[1]));
+        vst1q_s32(mtx_out + 2 * out_stride + 8, vreinterpretq_s32_u32(c2.val[2]));
+        vst1q_s32(mtx_out + 2 * out_stride + 12, vreinterpretq_s32_u32(c2.val[3]));
+        vst1q_s32(mtx_out + 3 * out_stride + 0, vreinterpretq_s32_u32(c3.val[0]));
+        vst1q_s32(mtx_out + 3 * out_stride + 4, vreinterpretq_s32_u32(c3.val[1]));
+        vst1q_s32(mtx_out + 3 * out_stride + 8, vreinterpretq_s32_u32(c3.val[2]));
+        vst1q_s32(mtx_out + 3 * out_stride + 12, vreinterpretq_s32_u32(c3.val[3]));
     },
     ina, inb, out);
 }
diff --git a/src/core/NEON/kernels/NEGEMMLowpFinalizeKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp
index 255e486365..bd550db54c 100644
--- a/src/core/NEON/kernels/NEGEMMLowpFinalizeKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "arm_compute/core/NEON/kernels/NEGEMMLowpFinalizeKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h"
 
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/Error.h"
@@ -44,15 +44,117 @@ namespace arm_compute
 class Coordinates;
 } // namespace arm_compute
 
-template <bool add_a_offset, bool add_b_offset>
-void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
+NEGEMMLowpOffsetContributionKernel::NEGEMMLowpOffsetContributionKernel()
+    : _vector_sum_col(nullptr), _vector_sum_row(nullptr), _mm_result(nullptr), _a_offset(0), _b_offset(0), _k_offset(0), _slide_vector_sum_col(true)
 {
-    const int32x4_t c_offset_s32 = vdupq_n_s32(_c_offset);
-    const int32x4_t shift_s32    = vdupq_n_s32(-_shift);
+}
+
+void NEGEMMLowpOffsetContributionKernel::configure(ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, int32_t k, int32_t a_offset, int32_t b_offset)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32);
+
+    // If a_offset == 0, vector_sum_col can be a nullptr
+    if(a_offset != 0)
+    {
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32);
+        ARM_COMPUTE_ERROR_ON(vector_sum_col->info()->dimension(0) != mm_result->info()->dimension(0));
+
+        TensorShape vector_sum_col_shape = vector_sum_col->info()->tensor_shape();
+        vector_sum_col_shape.collapse(1);
+
+        // Check if vector_sum_col_shape should be slidden or not
+        // Don't slide vector_sum_col_shape along the y dimension if vector_sum_col_shape has just 1 dimension and vector_sum_row_shape more than 1
+        // This scenario can happen when the the matrix multiplication is used to perform a convolution operation
+        _slide_vector_sum_col = vector_sum_col_shape[1] != 1;
+    }
+
+    // If b_offset == 0, vector_sum_row can be a nullptr
+    if(b_offset != 0)
+    {
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
+        ARM_COMPUTE_ERROR_ON(vector_sum_row->info()->dimension(0) != mm_result->info()->dimension(1));
+
+        TensorShape output_shape         = mm_result->info()->tensor_shape();
+        TensorShape vector_sum_row_shape = vector_sum_row->info()->tensor_shape();
+        vector_sum_row_shape.collapse(1);
+        output_shape.collapse(2);
+
+        ARM_COMPUTE_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[2], "mm_result tensor must have the same number of batches of output tensor");
+
+        if(a_offset != 0)
+        {
+            TensorShape vector_sum_col_shape = vector_sum_col->info()->tensor_shape();
+            vector_sum_col_shape.collapse(1);
+
+            ARM_COMPUTE_ERROR_ON_MSG(vector_sum_col_shape[1] != 1
+                                     && vector_sum_col_shape[1] != vector_sum_row_shape[1],
+                                     "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1");
+        }
+    }
+
+    _vector_sum_col = vector_sum_col;
+    _vector_sum_row = vector_sum_row;
+    _mm_result      = mm_result;
+    _a_offset       = a_offset;
+    _b_offset       = b_offset;
+    _k_offset       = a_offset * b_offset * k;
+
+    constexpr unsigned int num_elems_processed_per_iteration = 16;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*mm_result->info(), Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal mm_result_access(mm_result->info(), 0, num_elems_processed_per_iteration);
+
+    // Accordingly with a_offset and b_offset, we can have 4 cases:
+    // a_offset != 0 && b_offset != 0
+    // a_offset  = 0 && b_offset != 0
+    // a_offset != 0 && b_offset  = 0
+    // a_offset  = 0 && b_offset  = 0
+    if(a_offset != 0 && b_offset != 0)
+    {
+        AccessWindowStatic     vector_sum_row_access(vector_sum_row->info(), 0, 0, vector_sum_row->info()->dimension(0), 0);
+        AccessWindowHorizontal vector_sum_col_access(vector_sum_col->info(), 0, num_elems_processed_per_iteration);
+
+        update_window_and_padding(win,
+                                  vector_sum_col_access,
+                                  vector_sum_row_access,
+                                  mm_result_access);
+    }
+    else if(a_offset == 0 && b_offset != 0)
+    {
+        AccessWindowStatic vector_sum_row_access(vector_sum_row->info(), 0, 0, vector_sum_row->info()->dimension(0), 0);
+
+        update_window_and_padding(win,
+                                  vector_sum_row_access,
+                                  mm_result_access);
+    }
+    else if(a_offset != 0 && b_offset == 0)
+    {
+        AccessWindowHorizontal vector_sum_col_access(vector_sum_col->info(), 0, num_elems_processed_per_iteration);
+
+        update_window_and_padding(win,
+                                  vector_sum_col_access,
+                                  mm_result_access);
+    }
+    else
+    {
+        update_window_and_padding(win,
+                                  mm_result_access);
+    }
+
+    INEKernel::configure(win);
+}
+
+void NEGEMMLowpOffsetContributionKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
     Window collapsed_window = window.collapse_if_possible(IKernel::window(), Window::DimZ);
 
-    if(add_a_offset && add_b_offset) // true, true
+    if(_a_offset != 0 && _b_offset != 0) // true, true
     {
         // Set window for vector_sum_col
         Window win_vector_sum_col(collapsed_window);
@@ -70,7 +172,6 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
         Iterator vector_sum_col(_vector_sum_col, win_vector_sum_col);
         Iterator vector_sum_row(_vector_sum_row, win_vector_sum_row);
         Iterator mm_result(_mm_result, window);
-        Iterator out(_output, window);
 
         execute_window_loop(window, [&](const Coordinates & id)
         {
@@ -110,12 +211,6 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
             offset_term_s32.val[2] = vaddq_s32(offset_term_s32.val[2], vaddq_s32(a_offset_term_s32.val[2], b_offset_term_s32));
             offset_term_s32.val[3] = vaddq_s32(offset_term_s32.val[3], vaddq_s32(a_offset_term_s32.val[3], b_offset_term_s32));
 
-            // Add c_offset
-            offset_term_s32.val[0] = vaddq_s32(offset_term_s32.val[0], c_offset_s32);
-            offset_term_s32.val[1] = vaddq_s32(offset_term_s32.val[1], c_offset_s32);
-            offset_term_s32.val[2] = vaddq_s32(offset_term_s32.val[2], c_offset_s32);
-            offset_term_s32.val[3] = vaddq_s32(offset_term_s32.val[3], c_offset_s32);
-
             int32x4x4_t in_s32 =
             {
                 {
@@ -132,35 +227,15 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
             in_s32.val[2] = vaddq_s32(in_s32.val[2], offset_term_s32.val[2]);
             in_s32.val[3] = vaddq_s32(in_s32.val[3], offset_term_s32.val[3]);
 
-            // Multiply by c_mult_int
-            in_s32.val[0] = vmulq_n_s32(in_s32.val[0], _c_mult_int);
-            in_s32.val[1] = vmulq_n_s32(in_s32.val[1], _c_mult_int);
-            in_s32.val[2] = vmulq_n_s32(in_s32.val[2], _c_mult_int);
-            in_s32.val[3] = vmulq_n_s32(in_s32.val[3], _c_mult_int);
-
-            // Shift final result (negative value shift right)
-            in_s32.val[0] = vshlq_s32(in_s32.val[0], shift_s32);
-            in_s32.val[1] = vshlq_s32(in_s32.val[1], shift_s32);
-            in_s32.val[2] = vshlq_s32(in_s32.val[2], shift_s32);
-            in_s32.val[3] = vshlq_s32(in_s32.val[3], shift_s32);
-
-            // Convert S32 to U16
-            const int16x8x2_t in_s16 =
-            {
-                {
-                    vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
-                    vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])),
-                }
-            };
-
-            // Convert S16 to S8
-            const int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1]));
-
-            vst1q_s8(reinterpret_cast<int8_t *>(out.ptr()), out_s8);
+            // Store the result with the offset contribution
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 0, in_s32.val[0]);
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 4, in_s32.val[1]);
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 8, in_s32.val[2]);
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 12, in_s32.val[3]);
         },
-        vector_sum_col, vector_sum_row, mm_result, out);
+        vector_sum_col, vector_sum_row, mm_result);
     }
-    else if(!add_a_offset && add_b_offset) // false, true
+    else if((_a_offset == 0) && (_b_offset != 0)) // false, true
     {
         // Set window for vector_sum_row
         Window win_vector_sum_row(collapsed_window);
@@ -169,7 +244,6 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
 
         Iterator vector_sum_row(_vector_sum_row, win_vector_sum_row);
         Iterator mm_result(_mm_result, window);
-        Iterator out(_output, window);
 
         execute_window_loop(window, [&](const Coordinates & id)
         {
@@ -177,9 +251,6 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
             int32x4_t b_offset_term_s32 = vld1q_dup_s32(reinterpret_cast<const int32_t *>(vector_sum_row.ptr()) + id.y());
             b_offset_term_s32           = vmulq_n_s32(b_offset_term_s32, _b_offset);
 
-            // Add b_offset_term_s32 and c_offset_term_s32
-            int32x4_t offset_term_s32 = vaddq_s32(b_offset_term_s32, c_offset_s32);
-
             int32x4x4_t in_s32 =
             {
                 {
@@ -191,40 +262,20 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
             };
 
             // Add the offset terms to GEMM's result
-            in_s32.val[0] = vaddq_s32(in_s32.val[0], offset_term_s32);
-            in_s32.val[1] = vaddq_s32(in_s32.val[1], offset_term_s32);
-            in_s32.val[2] = vaddq_s32(in_s32.val[2], offset_term_s32);
-            in_s32.val[3] = vaddq_s32(in_s32.val[3], offset_term_s32);
-
-            // Multiply by c_mult_int
-            in_s32.val[0] = vmulq_n_s32(in_s32.val[0], _c_mult_int);
-            in_s32.val[1] = vmulq_n_s32(in_s32.val[1], _c_mult_int);
-            in_s32.val[2] = vmulq_n_s32(in_s32.val[2], _c_mult_int);
-            in_s32.val[3] = vmulq_n_s32(in_s32.val[3], _c_mult_int);
-
-            // Shift final result (negative value shift right)
-            in_s32.val[0] = vshlq_s32(in_s32.val[0], shift_s32);
-            in_s32.val[1] = vshlq_s32(in_s32.val[1], shift_s32);
-            in_s32.val[2] = vshlq_s32(in_s32.val[2], shift_s32);
-            in_s32.val[3] = vshlq_s32(in_s32.val[3], shift_s32);
-
-            // Convert S32 to U16
-            const int16x8x2_t in_s16 =
-            {
-                {
-                    vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
-                    vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])),
-                }
-            };
-
-            // Convert S16 to S8
-            const int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1]));
-
-            vst1q_s8(reinterpret_cast<int8_t *>(out.ptr()), out_s8);
+            in_s32.val[0] = vaddq_s32(in_s32.val[0], b_offset_term_s32);
+            in_s32.val[1] = vaddq_s32(in_s32.val[1], b_offset_term_s32);
+            in_s32.val[2] = vaddq_s32(in_s32.val[2], b_offset_term_s32);
+            in_s32.val[3] = vaddq_s32(in_s32.val[3], b_offset_term_s32);
+
+            // Store the result with the offset contribution
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 0, in_s32.val[0]);
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 4, in_s32.val[1]);
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 8, in_s32.val[2]);
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 12, in_s32.val[3]);
         },
-        vector_sum_row, mm_result, out);
+        vector_sum_row, mm_result);
     }
-    else if(add_a_offset && !add_b_offset) // true, false
+    else if((_a_offset != 0) && (_b_offset == 0)) // true, false
     {
         // Set window for vector_sum_col
         Window win_vector_sum_col(collapsed_window);
@@ -236,7 +287,6 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
 
         Iterator vector_sum_col(_vector_sum_col, win_vector_sum_col);
         Iterator mm_result(_mm_result, window);
-        Iterator out(_output, window);
 
         execute_window_loop(window, [&](const Coordinates & id)
         {
@@ -256,17 +306,6 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
             a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], _a_offset);
             a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], _a_offset);
 
-            // Add a_offset_term_s32 and b_offset_term_s32
-            int32x4x4_t offset_term_s32 =
-            {
-                {
-                    vaddq_s32(c_offset_s32, a_offset_term_s32.val[0]),
-                    vaddq_s32(c_offset_s32, a_offset_term_s32.val[1]),
-                    vaddq_s32(c_offset_s32, a_offset_term_s32.val[2]),
-                    vaddq_s32(c_offset_s32, a_offset_term_s32.val[3])
-                }
-            };
-
             int32x4x4_t in_s32 =
             {
                 {
@@ -278,232 +317,22 @@ void NEGEMMLowpFinalizeKernel::finalize(const Window &window)
             };
 
             // Add the offset terms to GEMM's result
-            in_s32.val[0] = vaddq_s32(in_s32.val[0], offset_term_s32.val[0]);
-            in_s32.val[1] = vaddq_s32(in_s32.val[1], offset_term_s32.val[1]);
-            in_s32.val[2] = vaddq_s32(in_s32.val[2], offset_term_s32.val[2]);
-            in_s32.val[3] = vaddq_s32(in_s32.val[3], offset_term_s32.val[3]);
-
-            // Multiply by c_mult_int
-            in_s32.val[0] = vmulq_n_s32(in_s32.val[0], _c_mult_int);
-            in_s32.val[1] = vmulq_n_s32(in_s32.val[1], _c_mult_int);
-            in_s32.val[2] = vmulq_n_s32(in_s32.val[2], _c_mult_int);
-            in_s32.val[3] = vmulq_n_s32(in_s32.val[3], _c_mult_int);
-
-            // Shift final result (negative value shift right)
-            in_s32.val[0] = vshlq_s32(in_s32.val[0], shift_s32);
-            in_s32.val[1] = vshlq_s32(in_s32.val[1], shift_s32);
-            in_s32.val[2] = vshlq_s32(in_s32.val[2], shift_s32);
-            in_s32.val[3] = vshlq_s32(in_s32.val[3], shift_s32);
-
-            // Convert S32 to S16
-            const int16x8x2_t in_s16 =
-            {
-                {
-                    vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
-                    vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))
-                }
-            };
-
-            // Convert S16 to S8
-            const int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1]));
-
-            vst1q_s8(reinterpret_cast<int8_t *>(out.ptr()), out_s8);
+            in_s32.val[0] = vaddq_s32(in_s32.val[0], a_offset_term_s32.val[0]);
+            in_s32.val[1] = vaddq_s32(in_s32.val[1], a_offset_term_s32.val[1]);
+            in_s32.val[2] = vaddq_s32(in_s32.val[2], a_offset_term_s32.val[2]);
+            in_s32.val[3] = vaddq_s32(in_s32.val[3], a_offset_term_s32.val[3]);
+
+            // Store the result with the offset contribution
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 0, in_s32.val[0]);
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 4, in_s32.val[1]);
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 8, in_s32.val[2]);
+            vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 12, in_s32.val[3]);
         },
-        vector_sum_col, mm_result, out);
+        vector_sum_col, mm_result);
     }
     else // false, false
     {
-        Iterator mm_result(_mm_result, window);
-        Iterator out(_output, window);
-
-        execute_window_loop(window, [&](const Coordinates & id)
-        {
-            int32x4x4_t in_s32 =
-            {
-                {
-                    vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 0),
-                    vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 4),
-                    vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 8),
-                    vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 12)
-                }
-            };
-
-            // Add the offset terms to GEMM's result
-            in_s32.val[0] = vaddq_s32(in_s32.val[0], c_offset_s32);
-            in_s32.val[1] = vaddq_s32(in_s32.val[1], c_offset_s32);
-            in_s32.val[2] = vaddq_s32(in_s32.val[2], c_offset_s32);
-            in_s32.val[3] = vaddq_s32(in_s32.val[3], c_offset_s32);
-
-            // Multiply by c_mult_int
-            in_s32.val[0] = vmulq_n_s32(in_s32.val[0], _c_mult_int);
-            in_s32.val[1] = vmulq_n_s32(in_s32.val[1], _c_mult_int);
-            in_s32.val[2] = vmulq_n_s32(in_s32.val[2], _c_mult_int);
-            in_s32.val[3] = vmulq_n_s32(in_s32.val[3], _c_mult_int);
-
-            // Shift final result (negative value shift right)
-            in_s32.val[0] = vshlq_s32(in_s32.val[0], shift_s32);
-            in_s32.val[1] = vshlq_s32(in_s32.val[1], shift_s32);
-            in_s32.val[2] = vshlq_s32(in_s32.val[2], shift_s32);
-            in_s32.val[3] = vshlq_s32(in_s32.val[3], shift_s32);
-
-            // Convert S32 to S16
-            const int16x8x2_t in_s16 =
-            {
-                {
-                    vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
-                    vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))
-                }
-            };
-
-            // Convert U16 to S8
-            const int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1]));
-
-            vst1q_s8(reinterpret_cast<int8_t *>(out.ptr()), out_s8);
-        },
-        mm_result, out);
+        // No offset contribution from matrix A and matrix B
+        return;
     }
 }
-
-NEGEMMLowpFinalizeKernel::NEGEMMLowpFinalizeKernel()
-    : _func(nullptr), _vector_sum_col(nullptr), _vector_sum_row(nullptr), _mm_result(nullptr), _output(nullptr), _a_offset(0), _b_offset(0), _c_offset(0), _k_offset(0), _c_mult_int(0), _shift(0),
-      _slide_vector_sum_col(true)
-{
-}
-
-void NEGEMMLowpFinalizeKernel::configure(const ITensor *vector_sum_col, const ITensor *vector_sum_row, const ITensor *mm_result, ITensor *output, int32_t num_mtx_a_cols, int32_t a_offset,
-                                         int32_t b_offset,
-                                         int32_t c_offset, int32_t c_mult_int, int32_t shift)
-{
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S8);
-
-    TensorShape mm_result_shape = mm_result->info()->tensor_shape();
-    TensorShape output_shape    = output->info()->tensor_shape();
-
-    mm_result_shape.collapse(2);
-    output_shape.collapse(2);
-
-    ARM_COMPUTE_ERROR_ON_MSG(mm_result_shape[2] != output_shape[2], "mm_result tensor must have the same number of batches of output tensor");
-
-    // If a_offset == 0, vector_sum_col can be a nullptr
-    if(a_offset != 0)
-    {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32);
-        ARM_COMPUTE_ERROR_ON(vector_sum_col->info()->dimension(0) != mm_result->info()->dimension(0));
-
-        TensorShape vector_sum_col_shape = vector_sum_col->info()->tensor_shape();
-        vector_sum_col_shape.collapse(1);
-
-        // Check if vector_sum_col_shape should be slidden or not
-        // Don't slide vector_sum_col_shape along the y dimension if vector_sum_col_shape has just 1 dimension and vector_sum_row_shape more than 1
-        // This scenario can happen when the the matrix multiplication is used to perform a convolution operation
-        _slide_vector_sum_col = vector_sum_col_shape[1] != 1;
-    }
-
-    // If b_offset == 0, vector_sum_row can be a nullptr
-    if(b_offset != 0)
-    {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
-        ARM_COMPUTE_ERROR_ON(vector_sum_row->info()->dimension(0) != mm_result->info()->dimension(1));
-
-        TensorShape vector_sum_row_shape = vector_sum_row->info()->tensor_shape();
-        vector_sum_row_shape.collapse(1);
-
-        ARM_COMPUTE_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[2], "mm_result tensor must have the same number of batches of output tensor");
-
-        if(a_offset != 0)
-        {
-            TensorShape vector_sum_col_shape = vector_sum_col->info()->tensor_shape();
-            vector_sum_col_shape.collapse(1);
-
-            ARM_COMPUTE_ERROR_ON_MSG(vector_sum_col_shape[1] != 1
-                                     && vector_sum_col_shape[1] != vector_sum_row_shape[1],
-                                     "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1");
-        }
-    }
-
-    _vector_sum_col = vector_sum_col;
-    _vector_sum_row = vector_sum_row;
-    _mm_result      = mm_result;
-    _output         = output;
-    _a_offset       = a_offset;
-    _b_offset       = b_offset;
-    _k_offset       = a_offset * b_offset * num_mtx_a_cols;
-    _c_offset       = c_offset;
-    _c_mult_int     = c_mult_int;
-    _shift          = shift;
-
-    constexpr unsigned int num_elems_processed_per_iteration = 16;
-
-    // Configure kernel window
-    Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
-
-    AccessWindowHorizontal mm_result_access(mm_result->info(), 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_result_access(output->info(), 0, num_elems_processed_per_iteration);
-
-    // Accordingly with a_offset and b_offset, we can have 4 cases:
-    // a_offset != 0 && b_offset != 0
-    // a_offset  = 0 && b_offset != 0
-    // a_offset != 0 && b_offset  = 0
-    // a_offset  = 0 && b_offset  = 0
-    if(a_offset != 0 && b_offset != 0)
-    {
-        // Set the function to use
-        _func = &NEGEMMLowpFinalizeKernel::finalize<true, true>;
-
-        AccessWindowStatic     vector_sum_row_access(vector_sum_row->info(), 0, 0, vector_sum_row->info()->dimension(0), 0);
-        AccessWindowHorizontal vector_sum_col_access(vector_sum_col->info(), 0, num_elems_processed_per_iteration);
-
-        update_window_and_padding(win,
-                                  vector_sum_col_access,
-                                  vector_sum_row_access,
-                                  mm_result_access,
-                                  output_result_access);
-    }
-    else if(a_offset == 0 && b_offset != 0)
-    {
-        // Set the function to use
-        _func = &NEGEMMLowpFinalizeKernel::finalize<false, true>;
-
-        AccessWindowStatic vector_sum_row_access(vector_sum_row->info(), 0, 0, vector_sum_row->info()->dimension(0), 0);
-
-        update_window_and_padding(win,
-                                  vector_sum_row_access,
-                                  mm_result_access,
-                                  output_result_access);
-    }
-    else if(a_offset != 0 && b_offset == 0)
-    {
-        // Set the function to use
-        _func = &NEGEMMLowpFinalizeKernel::finalize<true, false>;
-
-        AccessWindowHorizontal vector_sum_col_access(vector_sum_col->info(), 0, num_elems_processed_per_iteration);
-
-        update_window_and_padding(win,
-                                  vector_sum_col_access,
-                                  mm_result_access,
-                                  output_result_access);
-    }
-    else
-    {
-        // Set the function to use
-        _func = &NEGEMMLowpFinalizeKernel::finalize<false, false>;
-
-        update_window_and_padding(win,
-                                  mm_result_access,
-                                  output_result_access);
-    }
-
-    output_result_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
-
-    INEKernel::configure(win);
-}
-
-void NEGEMMLowpFinalizeKernel::run(const Window &window, const ThreadInfo &info)
-{
-    ARM_COMPUTE_UNUSED(info);
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
-
-    (this->*_func)(window);
-}
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
new file mode 100644
index 0000000000..aa3c280788
--- /dev/null
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <arm_neon.h>
+#include <cstddef>
+#include <cstdint>
+
+using namespace arm_compute;
+
+namespace arm_compute
+{
+class Coordinates;
+} // namespace arm_compute
+
+NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel()
+    : _input(nullptr), _output(nullptr), _result_offset(0), _result_mult_int(0), _result_shift(0)
+{
+}
+
+void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ITensor *input, ITensor *output, int result_offset, int result_mult_int, int result_shift)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+
+    _input           = input;
+    _output          = output;
+    _result_offset   = result_offset;
+    _result_mult_int = result_mult_int;
+    _result_shift    = result_shift;
+
+    constexpr unsigned int num_elems_processed_per_iteration = 16;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_result_access(output->info(), 0, num_elems_processed_per_iteration);
+
+    update_window_and_padding(win,
+                              input_access,
+                              output_result_access);
+
+    output_result_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    const int32x4_t result_offset_s32 = vdupq_n_s32(_result_offset);
+    const int32x4_t result_shift_s32  = vdupq_n_s32(-_result_shift);
+    const int32x4_t zero_s32          = vdupq_n_s32(0);
+
+    Iterator in(_input, window);
+    Iterator out(_output, window);
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        int32x4x4_t in_s32 =
+        {
+            {
+                vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 0),
+                vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 4),
+                vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 8),
+                vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 12)
+            }
+        };
+
+        // Add the offset terms to GEMM's result
+        in_s32.val[0] = vaddq_s32(in_s32.val[0], result_offset_s32);
+        in_s32.val[1] = vaddq_s32(in_s32.val[1], result_offset_s32);
+        in_s32.val[2] = vaddq_s32(in_s32.val[2], result_offset_s32);
+        in_s32.val[3] = vaddq_s32(in_s32.val[3], result_offset_s32);
+
+        // Multiply by c_mult_int
+        in_s32.val[0] = vmulq_n_s32(in_s32.val[0], _result_mult_int);
+        in_s32.val[1] = vmulq_n_s32(in_s32.val[1], _result_mult_int);
+        in_s32.val[2] = vmulq_n_s32(in_s32.val[2], _result_mult_int);
+        in_s32.val[3] = vmulq_n_s32(in_s32.val[3], _result_mult_int);
+
+        // Shift final result (negative value shift right)
+        in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32);
+        in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32);
+        in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32);
+        in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32);
+
+        // Saturate negative values
+        in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32);
+        in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32);
+        in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32);
+        in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32);
+
+        // Convert S32 to S16
+        const int16x8x2_t in_s16 =
+        {
+            {
+                vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
+                vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))
+            }
+        };
+
+        // Convert S16 to U8
+        const uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1]));
+
+        vst1q_u8(out.ptr(), out_u8);
+    },
+    in, out);
+}
\ No newline at end of file
diff --git a/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp
index 9df13ce0e3..81d9b5bb81 100644
--- a/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp
@@ -49,12 +49,12 @@ INEGEMMLowpReductionKernel::INEGEMMLowpReductionKernel()
 {
 }
 
-void NEGEMMLowpMatrixAReductionKernel::configure(const ITensor *mtx_a_interleaved4x4, ITensor *vector_sum_row, int32_t num_mtx_a_cols, bool is_interleaved4x4)
+void NEGEMMLowpMatrixAReductionKernel::configure(const ITensor *mtx_a, ITensor *vector_sum_row, int32_t num_mtx_a_cols, bool is_interleaved4x4)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_a_interleaved4x4, 1, DataType::S8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_a, 1, DataType::QASYMM8);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
 
-    _input       = mtx_a_interleaved4x4;
+    _input       = mtx_a;
     _output      = vector_sum_row;
     _k           = num_mtx_a_cols;
     _is_reshaped = is_interleaved4x4;
@@ -97,9 +97,9 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf
         execute_window_loop(collapsed_window, [&](const Coordinates & id)
         {
             // Note: Since the input is unsigned char, we can safely use unsigned int for the accumulation
-            int32x4_t sum_row = vdupq_n_s32(0);
+            uint32x4_t sum_row = vdupq_n_u32(0);
 
-            auto matrix_a = reinterpret_cast<const int8_t *>(in.ptr() + (id.x() / 4) * _input->info()->strides_in_bytes()[1] + id.y() * _input->info()->strides_in_bytes()[2]);
+            const uint8_t *matrix_a = (in.ptr() + (id.x() / 4) * _input->info()->strides_in_bytes()[1] + id.y() * _input->info()->strides_in_bytes()[2]);
 
 #if __arm__
             asm volatile("PLD [%0, #128*4]" ::"r"(matrix_a));
@@ -109,43 +109,43 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf
             // This for loop performs 4 accumulations
             for(; i <= (_k - 4); i += 4)
             {
-                const int8x16_t a0_s8 = vld1q_s8(matrix_a + i * 4);
+                const uint8x16_t a0_u8 = vld1q_u8(matrix_a + i * 4);
 
                 // Convert U8 to U16
-                int16x4x4_t a0_s16 =
+                uint16x4x4_t a0_u16 =
                 {
                     {
-                        vget_low_s16(vmovl_s8(vget_low_s8(a0_s8))),
-                        vget_high_s16(vmovl_s8(vget_low_s8(a0_s8))),
-                        vget_low_s16(vmovl_s8(vget_high_s8(a0_s8))),
-                        vget_high_s16(vmovl_s8(vget_high_s8(a0_s8)))
+                        vget_low_u16(vmovl_u8(vget_low_u8(a0_u8))),
+                        vget_high_u16(vmovl_u8(vget_low_u8(a0_u8))),
+                        vget_low_u16(vmovl_u8(vget_high_u8(a0_u8))),
+                        vget_high_u16(vmovl_u8(vget_high_u8(a0_u8)))
                     }
                 };
 
                 // Accumulate to U16
-                a0_s16.val[0] = vadd_s16(a0_s16.val[0], a0_s16.val[1]);
-                a0_s16.val[0] = vadd_s16(a0_s16.val[0], a0_s16.val[2]);
-                a0_s16.val[0] = vadd_s16(a0_s16.val[0], a0_s16.val[3]);
+                a0_u16.val[0] = vadd_u16(a0_u16.val[0], a0_u16.val[1]);
+                a0_u16.val[0] = vadd_u16(a0_u16.val[0], a0_u16.val[2]);
+                a0_u16.val[0] = vadd_u16(a0_u16.val[0], a0_u16.val[3]);
 
                 // Accumulate to U32
-                sum_row = vaddw_s16(sum_row, a0_s16.val[0]);
+                sum_row = vaddw_u16(sum_row, a0_u16.val[0]);
             }
 
             // This for loop performs the leftover accumulations
             for(; i < _k; ++i)
             {
-                const int8x8_t a0_s8 = vld1_s8(matrix_a + i * 4);
+                const uint8x8_t a0_u8 = vld1_u8(matrix_a + i * 4);
 
                 // Convert U8 to U16
-                const int16x4_t a0_s16 = vget_low_s16(vmovl_s8(a0_s8));
+                const uint16x4_t a0_u16 = vget_low_u16(vmovl_u8(a0_u8));
 
                 // Accumulate to U32
-                sum_row = vaddw_s16(sum_row, a0_s16);
+                sum_row = vaddw_u16(sum_row, a0_u16);
             }
 
             auto vector_sum_row = reinterpret_cast<int32_t *>(out.ptr());
 
-            vst1q_s32(vector_sum_row, sum_row);
+            vst1q_s32(vector_sum_row, vreinterpretq_s32_u32(sum_row));
         },
         in, out);
     }
@@ -154,10 +154,10 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf
         execute_window_loop(collapsed_window, [&](const Coordinates & id)
         {
             // Note: Since the input is unsigned char, we can safely use unsigned int for the accumulation
-            int32x4_t sum_row_s32 = vdupq_n_s32(0);
-            int32_t   sum_row     = 0;
+            uint32x4_t sum_row_u32 = vdupq_n_u32(0);
+            uint32_t   sum_row     = 0;
 
-            auto matrix_a = reinterpret_cast<const int8_t *>(in.ptr() + id.x() * _input->info()->strides_in_bytes()[1] + +id.y() * _input->info()->strides_in_bytes()[2]);
+            const uint8_t *matrix_a = (in.ptr() + id.x() * _input->info()->strides_in_bytes()[1] + +id.y() * _input->info()->strides_in_bytes()[2]);
 
 #if __arm__
             asm volatile("PLD [%0, #128*4]" ::"r"(matrix_a));
@@ -167,29 +167,29 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf
             // This for loop performs 16 accumulations
             for(; i <= (_k - 16); i += 16)
             {
-                const int8x16_t a0_s8 = vld1q_s8(matrix_a + i);
+                const uint8x16_t a0_u8 = vld1q_u8(matrix_a + i);
 
                 // Partial accumulations in U16
-                const int16x8_t tmp_sum0 = vaddl_s8(vget_low_s8(a0_s8), vget_high_s8(a0_s8));
+                const uint16x8_t tmp_sum0 = vaddl_u8(vget_low_u8(a0_u8), vget_high_u8(a0_u8));
 
                 // Accumulate to U32
-                sum_row_s32 = vaddq_s32(sum_row_s32, vpaddlq_s16(tmp_sum0));
+                sum_row_u32 = vaddq_u32(sum_row_u32, vpaddlq_u16(tmp_sum0));
             }
 
             // This for loop performs the leftover accumulations
             for(; i < _k; ++i)
             {
-                sum_row += static_cast<int32_t>(matrix_a[i]);
+                sum_row += static_cast<uint32_t>(matrix_a[i]);
             }
 
 #if defined(__aarch64__)
             // Reduction operation available on 64 bit architectures only
-            sum_row += vaddvq_s32(sum_row_s32);
+            sum_row += vaddvq_u32(sum_row_u32);
 #else  // __aarch64__
-            int32x2_t tmp = vpadd_s32(vget_high_s32(sum_row_s32), vget_low_s32(sum_row_s32));
-            tmp            = vpadd_s32(tmp, tmp);
+            uint32x2_t tmp = vpadd_u32(vget_high_u32(sum_row_u32), vget_low_u32(sum_row_u32));
+            tmp            = vpadd_u32(tmp, tmp);
 
-            sum_row += vget_lane_s32(tmp, 0);
+            sum_row += vget_lane_u32(tmp, 0);
 #endif // __aarch64__
 
             *(reinterpret_cast<int *>(out.ptr())) = static_cast<int>(sum_row);
@@ -198,12 +198,12 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf
     }
 }
 
-void NEGEMMLowpMatrixBReductionKernel::configure(const ITensor *mtx_b_transposed1xW, ITensor *vector_sum_col, int32_t num_mtx_b_rows, bool is_transposed1xW)
+void NEGEMMLowpMatrixBReductionKernel::configure(const ITensor *mtx_b, ITensor *vector_sum_col, int32_t num_mtx_b_rows, bool is_transposed1xW)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_b_transposed1xW, 1, DataType::S8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_b, 1, DataType::QASYMM8);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32);
 
-    _input       = mtx_b_transposed1xW;
+    _input       = mtx_b;
     _output      = vector_sum_col;
     _k           = num_mtx_b_rows;
     _is_reshaped = is_transposed1xW;
@@ -246,17 +246,17 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf
         execute_window_loop(collapsed_window, [&](const Coordinates & id)
         {
             // Note: Since the input is unsigned char, we can safely use unsigned int for the accumulation
-            int32x4x4_t sum_col =
+            uint32x4x4_t sum_col =
             {
                 {
-                    vdupq_n_s32(0),
-                    vdupq_n_s32(0),
-                    vdupq_n_s32(0),
-                    vdupq_n_s32(0)
+                    vdupq_n_u32(0),
+                    vdupq_n_u32(0),
+                    vdupq_n_u32(0),
+                    vdupq_n_u32(0)
                 }
             };
 
-            auto matrix_b = reinterpret_cast<const int8_t *>(in.ptr() + (id.x() / 16) * _input->info()->strides_in_bytes()[1] + id.y() * _input->info()->strides_in_bytes()[2]);
+            const uint8_t *matrix_b = in.ptr() + (id.x() / 16) * _input->info()->strides_in_bytes()[1] + id.y() * _input->info()->strides_in_bytes()[2];
 
 #if __arm__
             asm volatile("PLD [%0, #128*4]" ::"r"(matrix_b));
@@ -265,14 +265,14 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf
             int i = 0;
             for(; i < _k; ++i)
             {
-                const int8x16_t b0_s8 = vld1q_s8(matrix_b + i * 16);
+                const uint8x16_t b0_u8 = vld1q_u8(matrix_b + i * 16);
 
                 // Convert S8 to U16
-                const int16x8x2_t b0_s16 =
+                const uint16x8x2_t b0_u16 =
                 {
                     {
-                        vmovl_s8(vget_low_s8(b0_s8)),
-                        vmovl_s8(vget_high_s8(b0_s8))
+                        vmovl_u8(vget_low_u8(b0_u8)),
+                        vmovl_u8(vget_high_u8(b0_u8))
                     }
                 };
 
@@ -280,20 +280,20 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf
                 sum_col =
                 {
                     {
-                        vaddw_s16(sum_col.val[0], vget_low_s16(b0_s16.val[0])),
-                        vaddw_s16(sum_col.val[1], vget_high_s16(b0_s16.val[0])),
-                        vaddw_s16(sum_col.val[2], vget_low_s16(b0_s16.val[1])),
-                        vaddw_s16(sum_col.val[3], vget_high_s16(b0_s16.val[1]))
+                        vaddw_u16(sum_col.val[0], vget_low_u16(b0_u16.val[0])),
+                        vaddw_u16(sum_col.val[1], vget_high_u16(b0_u16.val[0])),
+                        vaddw_u16(sum_col.val[2], vget_low_u16(b0_u16.val[1])),
+                        vaddw_u16(sum_col.val[3], vget_high_u16(b0_u16.val[1]))
                     }
                 };
             }
 
             auto vector_sum_col = reinterpret_cast<int32_t *>(out.ptr());
 
-            vst1q_s32(vector_sum_col + 0, sum_col.val[0]);
-            vst1q_s32(vector_sum_col + 4, sum_col.val[1]);
-            vst1q_s32(vector_sum_col + 8, sum_col.val[2]);
-            vst1q_s32(vector_sum_col + 12, sum_col.val[3]);
+            vst1q_s32(vector_sum_col + 0, vreinterpretq_s32_u32(sum_col.val[0]));
+            vst1q_s32(vector_sum_col + 4, vreinterpretq_s32_u32(sum_col.val[1]));
+            vst1q_s32(vector_sum_col + 8, vreinterpretq_s32_u32(sum_col.val[2]));
+            vst1q_s32(vector_sum_col + 12, vreinterpretq_s32_u32(sum_col.val[3]));
         },
         in, out);
     }
@@ -326,17 +326,17 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf
             }
 
             // Note: Since the input is unsigned char, we can safely use unsigned int for the accumulation
-            int32x4x4_t sum_col =
+            uint32x4x4_t sum_col =
             {
                 {
-                    vdupq_n_s32(0),
-                    vdupq_n_s32(0),
-                    vdupq_n_s32(0),
-                    vdupq_n_s32(0)
+                    vdupq_n_u32(0),
+                    vdupq_n_u32(0),
+                    vdupq_n_u32(0),
+                    vdupq_n_u32(0)
                 }
             };
 
-            auto matrix_b = reinterpret_cast<const int8_t *>(inb.ptr() + id.y() * _input->info()->strides_in_bytes()[2]);
+            const uint8_t *matrix_b = inb.ptr() + id.y() * _input->info()->strides_in_bytes()[2];
 
 #if __arm__
             asm volatile("PLD [%0, #128*4]" ::"r"(matrix_b));
@@ -347,10 +347,10 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf
             // This for loop performs 4 accumulations
             for(; i <= (_k - 4); i += 4)
             {
-                const int8x16_t b0_s8 = vld1q_s8(matrix_b + 0 * in_b_stride);
-                const int8x16_t b1_s8 = vld1q_s8(matrix_b + 1 * in_b_stride);
-                const int8x16_t b2_s8 = vld1q_s8(matrix_b + 2 * in_b_stride);
-                const int8x16_t b3_s8 = vld1q_s8(matrix_b + 3 * in_b_stride);
+                const uint8x16_t b0_u8 = vld1q_u8(matrix_b + 0 * in_b_stride);
+                const uint8x16_t b1_u8 = vld1q_u8(matrix_b + 1 * in_b_stride);
+                const uint8x16_t b2_u8 = vld1q_u8(matrix_b + 2 * in_b_stride);
+                const uint8x16_t b3_u8 = vld1q_u8(matrix_b + 3 * in_b_stride);
 
 #if __arm__
                 asm volatile("PLD [%0, #128*1]" ::"r"(matrix_b + 1 * in_b_stride));
@@ -360,31 +360,31 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf
 #endif /* __arm__ */
 
                 // Partial accumulation in u16
-                int16x8x2_t tmp_sum =
+                uint16x8x2_t tmp_sum =
                 {
                     {
-                        vdupq_n_s16(0),
-                        vdupq_n_s16(0)
+                        vdupq_n_u16(0),
+                        vdupq_n_u16(0)
                     }
                 };
 
-                tmp_sum.val[0] = vaddw_s8(tmp_sum.val[0], vget_low_s8(b0_s8));
-                tmp_sum.val[0] = vaddw_s8(tmp_sum.val[0], vget_low_s8(b1_s8));
-                tmp_sum.val[0] = vaddw_s8(tmp_sum.val[0], vget_low_s8(b2_s8));
-                tmp_sum.val[0] = vaddw_s8(tmp_sum.val[0], vget_low_s8(b3_s8));
-                tmp_sum.val[1] = vaddw_s8(tmp_sum.val[1], vget_high_s8(b0_s8));
-                tmp_sum.val[1] = vaddw_s8(tmp_sum.val[1], vget_high_s8(b1_s8));
-                tmp_sum.val[1] = vaddw_s8(tmp_sum.val[1], vget_high_s8(b2_s8));
-                tmp_sum.val[1] = vaddw_s8(tmp_sum.val[1], vget_high_s8(b3_s8));
+                tmp_sum.val[0] = vaddw_u8(tmp_sum.val[0], vget_low_u8(b0_u8));
+                tmp_sum.val[0] = vaddw_u8(tmp_sum.val[0], vget_low_u8(b1_u8));
+                tmp_sum.val[0] = vaddw_u8(tmp_sum.val[0], vget_low_u8(b2_u8));
+                tmp_sum.val[0] = vaddw_u8(tmp_sum.val[0], vget_low_u8(b3_u8));
+                tmp_sum.val[1] = vaddw_u8(tmp_sum.val[1], vget_high_u8(b0_u8));
+                tmp_sum.val[1] = vaddw_u8(tmp_sum.val[1], vget_high_u8(b1_u8));
+                tmp_sum.val[1] = vaddw_u8(tmp_sum.val[1], vget_high_u8(b2_u8));
+                tmp_sum.val[1] = vaddw_u8(tmp_sum.val[1], vget_high_u8(b3_u8));
 
                 // Accumulate to U32
                 sum_col =
                 {
                     {
-                        vaddw_s16(sum_col.val[0], vget_low_s16(tmp_sum.val[0])),
-                        vaddw_s16(sum_col.val[1], vget_high_s16(tmp_sum.val[0])),
-                        vaddw_s16(sum_col.val[2], vget_low_s16(tmp_sum.val[1])),
-                        vaddw_s16(sum_col.val[3], vget_high_s16(tmp_sum.val[1]))
+                        vaddw_u16(sum_col.val[0], vget_low_u16(tmp_sum.val[0])),
+                        vaddw_u16(sum_col.val[1], vget_high_u16(tmp_sum.val[0])),
+                        vaddw_u16(sum_col.val[2], vget_low_u16(tmp_sum.val[1])),
+                        vaddw_u16(sum_col.val[3], vget_high_u16(tmp_sum.val[1]))
                     }
                 };
 
@@ -394,14 +394,14 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf
             // This for loop perfoms the leftover accumulations
             for(; i < _k; ++i)
             {
-                const int8x16_t b0_s8 = vld1q_s8(matrix_b + 0 * in_b_stride);
+                const uint8x16_t b0_u8 = vld1q_u8(matrix_b + 0 * in_b_stride);
 
                 // Convert S8 to S16
-                const int16x8x2_t b0_s16 =
+                const uint16x8x2_t b0_u16 =
                 {
                     {
-                        vmovl_s8(vget_low_s8(b0_s8)),
-                        vmovl_s8(vget_high_s8(b0_s8))
+                        vmovl_u8(vget_low_u8(b0_u8)),
+                        vmovl_u8(vget_high_u8(b0_u8))
                     }
                 };
 
@@ -409,10 +409,10 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf
                 sum_col =
                 {
                     {
-                        vaddw_s16(sum_col.val[0], vget_low_s16(b0_s16.val[0])),
-                        vaddw_s16(sum_col.val[1], vget_high_s16(b0_s16.val[0])),
-                        vaddw_s16(sum_col.val[2], vget_low_s16(b0_s16.val[1])),
-                        vaddw_s16(sum_col.val[3], vget_high_s16(b0_s16.val[1]))
+                        vaddw_u16(sum_col.val[0], vget_low_u16(b0_u16.val[0])),
+                        vaddw_u16(sum_col.val[1], vget_high_u16(b0_u16.val[0])),
+                        vaddw_u16(sum_col.val[2], vget_low_u16(b0_u16.val[1])),
+                        vaddw_u16(sum_col.val[3], vget_high_u16(b0_u16.val[1]))
                     }
                 };
 
@@ -421,10 +421,10 @@ void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInf
 
             auto vector_sum_col = reinterpret_cast<int32_t *>(out.ptr());
 
-            vst1q_s32(vector_sum_col + 0, sum_col.val[0]);
-            vst1q_s32(vector_sum_col + 4, sum_col.val[1]);
-            vst1q_s32(vector_sum_col + 8, sum_col.val[2]);
-            vst1q_s32(vector_sum_col + 12, sum_col.val[3]);
+            vst1q_s32(vector_sum_col + 0, vreinterpretq_s32_u32(sum_col.val[0]));
+            vst1q_s32(vector_sum_col + 4, vreinterpretq_s32_u32(sum_col.val[1]));
+            vst1q_s32(vector_sum_col + 8, vreinterpretq_s32_u32(sum_col.val[2]));
+            vst1q_s32(vector_sum_col + 12, vreinterpretq_s32_u32(sum_col.val[3]));
         },
         inb, out);
     }
diff --git a/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp b/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp
index 7f4ee1ec49..7f83144e12 100644
--- a/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp
@@ -43,7 +43,8 @@ using namespace arm_compute;
 
 void NEGEMMTranspose1xWKernel::configure(const ITensor *input, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::U8, DataType::S8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16,
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::U8, DataType::S8, DataType::U16, DataType::S16, DataType::U32, DataType::S32,
+                                                  DataType::F16,
                                                   DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);