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>

10 files changed, 264 insertions, 228 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 8dedf38b3e..d78cec2a62 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -61,8 +61,9 @@
 #include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMInterleaveBlockedKernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMLowpFinalizeKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAdditionKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpFinalizeKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpFinalizeKernel.h
deleted file mode 100644
index 8908fabc1e..0000000000
--- a/arm_compute/core/NEON/kernels/NEGEMMLowpFinalizeKernel.h
+++ /dev/null
@@ -1,103 +0,0 @@
-/*
- * 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.
- */
-#ifndef __ARM_COMPUTE_NEGEMMLOWPFINALIZEKERNEL_H__
-#define __ARM_COMPUTE_NEGEMMLOWPFINALIZEKERNEL_H__
-
-#include "arm_compute/core/NEON/INEKernel.h"
-
-namespace arm_compute
-{
-class ITensor;
-
-/* NEON kernel used to finalize the GEMMLowp result
- *
- * This kernel performs the following computations:
- *
- *  -# Add offset terms to final result
- *  -# Multiply each entry of result and round to nearest integer
- *  -# Clamp the resulting int32 values to the [0..255] range and cast to uint8.
- */
-class NEGEMMLowpFinalizeKernel : public INEKernel
-{
-public:
-    /** Constructor */
-    NEGEMMLowpFinalizeKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpFinalizeKernel(const NEGEMMLowpFinalizeKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    NEGEMMLowpFinalizeKernel &operator=(const NEGEMMLowpFinalizeKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpFinalizeKernel(NEGEMMLowpFinalizeKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    NEGEMMLowpFinalizeKernel &operator=(NEGEMMLowpFinalizeKernel &&) = default;
-    /** Initialise the kernel's input and output.
-     *
-     * @note The input row-vectors  @p vector_sum_col and @p vector_sum_row must be the output of @ref NEGEMMLowpMatrixBReductionKernel and @ref NEGEMMLowpMatrixAReductionKernel kernels.
-     *       These 2 vectors are needed to handle the offset of matrix product
-     *       https://github.com/google/gemmlowp/blob/master/doc/low-precision.md
-     *
-     * @param[in]  vector_sum_col Input row-vector of sums of all the entries in each column of input1.
-     *                            Note: vector_sum_col can be a nullptr in case a_offset = 0. Data type supported: S32
-     * @param[in]  vector_sum_row Input row-vector of sums of all the entries in each row of input0.
-     *                            Note: vector_sum_row can be a nullptr in case b_offset = 0. Data type supported: same as @p vector_sum_col
-     * @param[in]  mm_result      Input tensor containing the result of @ref NEGEMMLowpMatrixMultiplyKernel. Data type supported: same as @p vector_sum_col
-     * @param[out] output         Output tensor containing the result of GEMMLowP. Data type supported: S8
-     * @param[in]  num_mtx_a_cols Number of matrix A columns
-     * @param[in]  a_offset       Offset to be added to each element of the matrix A.
-     * @param[in]  b_offset       Offset to be added to each element of the matrix B.
-     * @param[in]  c_offset       Offset to be added to each element of the output matrix
-     * @param[in]  c_mult_int     Value to be multiplied to each entry of the result.
-     * @param[in]  shift          Number of bits to shift right the result.
-     */
-    void 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);
-
-    // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
-
-private:
-    /** Template function to run the finalize kernel
-     *
-     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
-     */
-    template <bool add_a_offset, bool add_b_offset>
-    void finalize(const Window &window);
-    using FinalizeFunctionPtr = void (NEGEMMLowpFinalizeKernel::*)(const Window &window);
-
-    FinalizeFunctionPtr _func;
-    const ITensor      *_vector_sum_col;
-    const ITensor      *_vector_sum_row;
-    const ITensor      *_mm_result;
-    ITensor            *_output;
-    int32_t             _a_offset;
-    int32_t             _b_offset;
-    int32_t             _c_offset;
-    int32_t             _k_offset;
-    int32_t             _c_mult_int;
-    int32_t             _shift;
-    bool                _slide_vector_sum_col;
-};
-} // namespace arm_compute
-
-#endif /* __ARM_COMPUTE_NEGEMMLOWPFINALIZEKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h
index f145eb6ca3..e9bfe4ea07 100644
--- a/arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h
@@ -58,8 +58,8 @@ public:
      * The input matrices @p input0 and @p input1 must be the output of the kernels: @ref NEGEMMInterleave4x4Kernel and @ref NEGEMMTranspose1xWKernel. These two
      * kernels change the layout of the original matrices to be more cache-friendly.
      *
-     * @param[in]  input0 Input tensor containing the interleaved Matrix A. Data type supported: S8
-     * @param[in]  input1 Input tensor containing the transposed Matrix B. Data type supported: same as @p input0
+     * @param[in]  input0 Input tensor containing the interleaved Matrix A. Data type supported: ASYMM8
+     * @param[in]  input1 Input tensor containing the transposed1xW Matrix B. Data type supported: same as @p input0
      * @param[out] output Output tensor to store the result of matrix multiplication. Data type supported: S32
      */
     void configure(const ITensor *input0, const ITensor *input1, ITensor *output);
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h
new file mode 100644
index 0000000000..04b84339b0
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h
@@ -0,0 +1,79 @@
+/*
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONKERNEL_H__
+#define __ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/* NEON kernel used to add the offset contribution after @ref NEGEMMLowpMatrixMultiplyKernel. The computation is performed in-place
+ *
+ * This kernel takes a final int32 accumulator value (the output of @NEGEMMLowpMatrixMultiplyKernel),
+ * and adds to it the offset contribution of matrix A and matrix B in-place.
+ *
+ */
+class NEGEMMLowpOffsetContributionKernel : public INEKernel
+{
+public:
+    /** Constructor */
+    NEGEMMLowpOffsetContributionKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpOffsetContributionKernel(const NEGEMMLowpOffsetContributionKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpOffsetContributionKernel &operator=(const NEGEMMLowpOffsetContributionKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpOffsetContributionKernel(NEGEMMLowpOffsetContributionKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpOffsetContributionKernel &operator=(NEGEMMLowpOffsetContributionKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in, out] mm_result      Input tensor containing the result of @ref NEGEMMLowpMatrixMultiplyKernel. Data type supported: S32
+     * @param[in]      vector_sum_col Input row-vector of sums of all the entries in each column of matrix B.
+     *                                Note: vector_sum_col can be a nullptr in case a_offset = 0. Data type supported: same as @p mm_result
+     * @param[in]      vector_sum_row Input row-vector of sums of all the entries in each row of matrix A.
+     *                                Note: vector_sum_row can be a nullptr in case b_offset = 0. Data type supported: same as @p mm_result
+     * @param[in]      k              Number of matrix A columns or Matrix B rows
+     * @param[in]      a_offset       Offset to be added to each element of the matrix A.
+     * @param[in]      b_offset       Offset to be added to each element of the matrix B.
+     */
+    void 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);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_vector_sum_col;
+    const ITensor *_vector_sum_row;
+    ITensor       *_mm_result;
+    int32_t        _a_offset;
+    int32_t        _b_offset;
+    int32_t        _k_offset;
+    bool           _slide_vector_sum_col;
+};
+} // namespace arm_compute
+
+#endif /* __ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h
new file mode 100644
index 0000000000..65f1042b9c
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h
@@ -0,0 +1,78 @@
+/*
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALE_H__
+#define __ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALE_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/* NEON kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
+ * The following computations will be performed by the kernel:
+ *
+ *  -# Add offset terms to final result
+ *  -# Multiply each entry of result and round to nearest integer
+ *  -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
+ *
+ */
+class NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel : public INEKernel
+{
+public:
+    /** Constructor */
+    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel(NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &operator=(NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+    * @param[in]  input           Input tensor. Data type supported: S32
+    * @param[out] output          Output tensor. Data type supported: Data type supported: QASYMM8
+    * @param[in]  result_offset   Offset to be added to each element of the input matrix
+    * @param[in]  result_mult_int Value to be multiplied to each element of the input matrix when once the result_offset has been add
+    * @param[in]  result_shift    Number of bits to shift right the result before converting back to QASYMM8
+     */
+    void configure(const ITensor *input, ITensor *output, int result_offset, int result_mult_int, int result_shift);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_input;
+    ITensor       *_output;
+    int32_t        _result_offset;
+    int32_t        _result_mult_int;
+    int32_t        _result_shift;
+};
+} // namespace arm_compute
+
+#endif /* __ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALE_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h
index a069969681..6eee54a9f0 100644
--- a/arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h
@@ -45,10 +45,9 @@ public:
     /** Allow instances of this class to be moved */
     INEGEMMLowpReductionKernel &operator=(INEGEMMLowpReductionKernel &&) = default;
 
-public:
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  input       Input tensor containing the interleaved or transposed matrix. Data type supported: S8
+     * @param[in]  input       Input tensor. Data type supported: S8
      * @param[out] output      Output row-vector of sums of all the entries in each row/col of input tensor. Data type supported: S32
      * @param[in]  k           Number of matrix A columns (or matrix B rows)
      * @param[in]  is_reshaped True if the input tensor has been reshaped
@@ -72,14 +71,12 @@ class NEGEMMLowpMatrixAReductionKernel : public INEGEMMLowpReductionKernel
 public:
     /** Initialise the kernel's input and output.
      *
-     * @note The input matrix @p mtx_a_interleaved4x4 must be the output of @ref NEGEMMInterleave4x4Kernel.
-     *
-     * @param[in]  mtx_a_interleaved4x4 Input tensor containing the interleaved Matrix A. Data type supported: U8
-     * @param[out] vector_sum_row       Output row-vector of sums of all the entries in each row of mtx_a_interleaved4x4. Data type supported: S32
-     * @param[in]  num_mtx_a_cols       Number of matrix A columns
-     * @param[in]  is_interleaved4x4    True if the input tensor is interleaved4x4
+     * @param[in]  mtx_a             Input tensor. Data type supported: QASYMM8
+     * @param[out] vector_sum_row    Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32
+     * @param[in]  num_mtx_a_cols    Number of matrix A columns
+     * @param[in]  is_interleaved4x4 True if the matrix A has been interleaved4x4
      */
-    void configure(const ITensor *mtx_a_interleaved4x4, ITensor *vector_sum_row, int32_t num_mtx_a_cols, bool is_interleaved4x4) override;
+    void configure(const ITensor *mtx_a, ITensor *vector_sum_row, int32_t num_mtx_a_cols, bool is_interleaved4x4) override;
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
@@ -95,14 +92,12 @@ class NEGEMMLowpMatrixBReductionKernel : public INEGEMMLowpReductionKernel
 public:
     /** Initialise the kernel's input and output.
      *
-     * @note The input matrix @p mtx_b_transposed1xW must be the output of @ref NEGEMMTranspose1xWKernel kernel.
-     *
-     * @param[in]  mtx_b_transposed1xW Input tensor containing the transposed Matrix B. Data type supported: Data type supported: U8
-     * @param[out] vector_sum_col      Output row-vector of sums of all the entries in each column of mtx_b_transposed1xW. Data type supported: S32
-     * @param[in]  num_mtx_b_rows      Number of matrix B rows
-     * @param[in]  is_transposed1xW    True if the input tensor is transposed 1xW
+     * @param[in]  mtx_b            Input tensor. Data type supported: Data type supported: QASYMM8
+     * @param[out] vector_sum_col   Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32
+     * @param[in]  num_mtx_b_rows   Number of matrix B rows
+     * @param[in]  is_transposed1xW True if the input tensor is transposed 1xW
      */
-    void configure(const ITensor *mtx_b_transposed1xW, ITensor *vector_sum_col, int32_t num_mtx_b_rows, bool is_transposed1xW) override;
+    void configure(const ITensor *mtx_b, ITensor *vector_sum_col, int32_t num_mtx_b_rows, bool is_transposed1xW) override;
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 563ade288a..118603b20b 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -60,8 +60,8 @@
 #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMM.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMMInterleave4x4.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMLowp.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMMTranspose1xW.h"
 #include "arm_compute/runtime/NEON/functions/NEGaussian3x3.h"
 #include "arm_compute/runtime/NEON/functions/NEGaussian5x5.h"
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowp.h b/arm_compute/runtime/NEON/functions/NEGEMMLowp.h
deleted file mode 100644
index 59c919e161..0000000000
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowp.h
+++ /dev/null
@@ -1,94 +0,0 @@
-/*
- * 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.
- */
-#ifndef __ARM_COMPUTE_NEGEMMLOWP_H__
-#define __ARM_COMPUTE_NEGEMMLOWP_H__
-
-#include "arm_compute/core/NEON/INEKernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMLowpFinalizeKernel.h"
-#include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h"
-#include "arm_compute/runtime/IFunction.h"
-#include "arm_compute/runtime/IMemoryManager.h"
-#include "arm_compute/runtime/MemoryGroup.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
-#include "arm_compute/runtime/Tensor.h"
-
-#include <memory>
-
-namespace arm_compute
-{
-class ITensor;
-
-/** Basic function to execute GEMMLowp on NEON. This function calls the following NEON kernels/function:
- *
- *  -# @ref NEGEMMLowpMatrixAReductionKernel
- *  -# @ref NEGEMMLowpMatrixBReductionKernel
- *  -# @ref NEGEMMLowpMatrixMultiplyCore
- *  -# @ref NEGEMMLowpFinalizeKernel
- *
-*/
-class NEGEMMLowp : public IFunction
-{
-public:
-    /** Constructor */
-    NEGEMMLowp(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
-    /** Initialise the kernel's inputs, output
-    *
-    * @note GEMM_LOWP:  low precision GEMM kernel
-    *  This kernel performs the following computations:
-    *
-    *  -# Convert a values from int8 to int32 and add a_offset to each of them.
-    *  -# Convert b values from int8 to int32 and add b_offset to each of them.
-    *  -# Compute the int32 matrix product of the resulting a * b.
-    *  -# Add output_offset to each entry of the result.
-    *  -# Multiply each entry of the result and round to the nearest integer
-    *  -# Clamp the resulting int32 values to the [0..255] range and cast to int8.
-    *
-    * @param[in]  a               First input tensor  (Matrix A). Data type supported: S8.
-    * @param[in]  b               Second input tensor (Matrix B). Data type supported: same as @p a
-    * @param[out] output          Output tensor. Data type supported: same as @p a.
-    * @param[in]  a_offset        Offset to be added to each element of the matrix A.
-    * @param[in]  b_offset        Offset to be added to each element of the matrix B.
-    * @param[in]  c_offset        Offset to be added to each element of the output matrix
-    * @param[in]  output_mult_int Value to be multiplied to each element of the output matrix
-    * @param[in]  shift           Number of bits to shift right the result.
-    */
-    void configure(const ITensor *a, const ITensor *b, ITensor *output, int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t output_mult_int, int32_t shift);
-
-    // Inherited methods overridden:
-    void run() override;
-
-private:
-    MemoryGroup                      _memory_group;
-    NEGEMMLowpMatrixMultiplyCore     _mm_func;
-    NEGEMMLowpMatrixAReductionKernel _mtx_a_reduction_kernel;
-    NEGEMMLowpMatrixBReductionKernel _mtx_b_reduction_kernel;
-    NEGEMMLowpFinalizeKernel         _finalize_kernel;
-    Tensor                           _vector_sum_col;
-    Tensor                           _vector_sum_row;
-    Tensor                           _mm_output;
-    int32_t                          _a_offset;
-    int32_t                          _b_offset;
-};
-}
-#endif /*__ARM_COMPUTE_NEGEMMLOWP_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
index c81a432295..0c441df4b9 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
@@ -25,6 +25,8 @@
 #define __ARM_COMPUTE_NEGEMMLOWPMATRIXMULTIPLYCORE_H__
 
 #include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
 #include "arm_compute/runtime/MemoryGroup.h"
@@ -41,11 +43,13 @@ class ITensor;
  *  -# @ref NEGEMMInterleave4x4Kernel
  *  -# @ref NEGEMMTranspose1xWKernel
  *  -# @ref NEGEMMLowpMatrixMultiplyKernel
+ *  -# @ref NEGEMMLowpOffsetContributionKernel
  *
  * otherwise if the DOT product instruction is available:
  *
  *  -# @ref NEGEMMInterleaveBlockedKernel
  *  -# @ref NEGEMMLowpAArch64V8P4Kernel
+ *  -# @ref NEGEMMLowpOffsetContributionKernel
  *
 */
 class NEGEMMLowpMatrixMultiplyCore : public IFunction
@@ -58,11 +62,11 @@ public:
     * @note GEMM_LOWP:  low precision GEMM kernel
     *  This kernel performs the following computations:
     *
-    *  -# Convert a values from uint8 to int32
-    *  -# Convert b values from uint8 to int32
-    *  -# Compute the int32 matrix product of the resulting a * b.
+    *  -# Convert a values from QASYMM8 to int32 and add a_offset to each of them.
+    *  -# Convert b values from QASYMM8 to int32 add b_offset to each of them.
+    *  -# Compute the matrix product of the resulting a * b in int32.
     *
-    * @param[in]  a      First input tensor  (Matrix A). Data type supported: U8.
+    * @param[in]  a      First input tensor  (Matrix A). Data type supported: QASYMM8.
     * @param[in]  b      Second input tensor (Matrix B). Data type supported: same as @p a
     * @param[out] output Output tensor. Data type supported: Data type supported: S32
     */
@@ -72,13 +76,20 @@ public:
     void run() override;
 
 private:
-    MemoryGroup                _memory_group;
-    std::unique_ptr<INEKernel> _mm_kernel;
-    std::unique_ptr<INEKernel> _mtx_a_reshape_kernel;
-    std::unique_ptr<INEKernel> _mtx_b_reshape_kernel;
-    Tensor                     _tmp_a;
-    Tensor                     _tmp_b;
-    Tensor                     _workspace;
+    MemoryGroup                        _memory_group;
+    std::unique_ptr<INEKernel>         _mm_kernel;
+    std::unique_ptr<INEKernel>         _mtx_a_reshape_kernel;
+    std::unique_ptr<INEKernel>         _mtx_b_reshape_kernel;
+    NEGEMMLowpMatrixAReductionKernel   _mtx_a_reduction_kernel;
+    NEGEMMLowpMatrixBReductionKernel   _mtx_b_reduction_kernel;
+    NEGEMMLowpOffsetContributionKernel _offset_contribution_kernel;
+    Tensor                             _vector_sum_col;
+    Tensor                             _vector_sum_row;
+    Tensor                             _tmp_a;
+    Tensor                             _tmp_b;
+    Tensor                             _workspace;
+    int32_t                            _a_offset;
+    int32_t                            _b_offset;
 };
 }
 #endif /*__ARM_COMPUTE_NEGEMMLOWPMATRIXMULTIPLYCORE_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
new file mode 100644
index 0000000000..8557ef42e1
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
@@ -0,0 +1,69 @@
+/*
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__
+#define __ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__
+
+#include "arm_compute/runtime/NEON/INESimpleFunction.h"
+
+/** This file contains all available output stages for GEMMLowp on NEON.
+ *
+ *  In gemmlowp, the "output stage" is the process that takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyCore),
+ *  and processes it to obtain the final ASYMM8 value.
+ *
+ *  More information about the GEMMLowp output stage can be found at https://github.com/google/gemmlowp/blob/master/doc/output.md
+ */
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Basic function to execute NEGEMMLowpQuantizeDownInt32ToUint8Scale on NEON.
+ *
+ *  NEGEMMLowpQuantizeDownInt32ToUint8Scale depends on 3 parameters: result_offset, result_mult_int, result_shift
+ *  The final result is:
+ *
+ *  ((input[i][k] + result_offset) * result_mult_int + rounding) >> result_shift
+ *
+ *  where rounding = (result_shift < 1) ? 0 : (1 << (result_shift - 1))
+ *
+ *  This function calls the following NEON kernels:
+ *
+ * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel
+ *
+*/
+class NEGEMMLowpQuantizeDownInt32ToUint8Scale : public INESimpleFunction
+{
+public:
+    /** Initialise the kernel's inputs, output
+    *
+    * @param[in]  input           Input tensor. It is the output of @ref NEGEMMLowpMatrixMultiplyCore function. Data type supported: S32
+    * @param[out] output          Output tensor. Data type supported: Data type supported: QASYMM8
+    * @param[in]  result_offset   Offset to be added to each element of the input matrix
+    * @param[in]  result_mult_int Value to be multiplied to each element of the input matrix when once the result_offset has been add
+    * @param[in]  result_shift    Number of bits to shift right the result before converting back to QASYMM8
+    */
+    void configure(const ITensor *input, ITensor *output, int result_offset, int result_mult_int, int result_shift);
+};
+}
+#endif /*__ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__ */
\ No newline at end of file