COMPMID-2847: Fuse output stage in GEMMLowpMatrixMultiplyReshapedOnlyRHS

Change-Id: Icd60eb368a34295434e8c141885b4666973a92a1
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2732
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 514 insertions, 7 deletions

diff --git a/src/core/CL/cl_kernels/gemm_helpers.h b/src/core/CL/cl_kernels/gemm_helpers.h
index 66e83c3558..79a9f094df 100644
--- a/src/core/CL/cl_kernels/gemm_helpers.h
+++ b/src/core/CL/cl_kernels/gemm_helpers.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -140,6 +140,118 @@
 #define LOAD_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) LOAD_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)
 /** @} */ // end of group LOAD_BLOCK
 
+/** Loads the elements from 0 to n-1 in the given variables (BASENAME0 to BASENAMEn-1).
+ * @name LOAD_ELEMENT_n
+ *
+ * @param[in] N0        The number of rows to load
+ * @param[in] DATA_TYPE The data type of variables
+ * @param[in] BASENAME  The basename of the destination variables for the loaded rows
+ * @param[in] PTR       The base pointer
+ * @param[in] OFFSET    The offset within a row
+ * @param[in] STRIDE_Y  The stride value in y-axis direction
+ * @{
+ */
+#define LOAD_ELEMENT_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##0 = *((__global DATA_TYPE *)(PTR + OFFSET + 0 * STRIDE_Y));
+
+#define LOAD_ELEMENT_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##1 = *((__global DATA_TYPE *)(PTR + OFFSET + 1 * STRIDE_Y));
+
+#define LOAD_ELEMENT_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##2 = *((__global DATA_TYPE *)(PTR + OFFSET + 2 * STRIDE_Y));
+
+#define LOAD_ELEMENT_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##3 = *((__global DATA_TYPE *)(PTR + OFFSET + 3 * STRIDE_Y));
+
+#define LOAD_ELEMENT_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##4 = *((__global DATA_TYPE *)(PTR + OFFSET + 4 * STRIDE_Y));
+
+#define LOAD_ELEMENT_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##5 = *((__global DATA_TYPE *)(PTR + OFFSET + 5 * STRIDE_Y));
+
+#define LOAD_ELEMENT_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##6 = *((__global DATA_TYPE *)(PTR + OFFSET + 6 * STRIDE_Y));
+
+#define LOAD_ELEMENT_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##7 = *((__global DATA_TYPE *)(PTR + OFFSET + 7 * STRIDE_Y));
+
+#define LOAD_ELEMENT_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##8 = *((__global DATA_TYPE *)(PTR + OFFSET + 8 * STRIDE_Y));
+
+#define LOAD_ELEMENT_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)      \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##9 = *((__global DATA_TYPE *)(PTR + OFFSET + 9 * STRIDE_Y));
+
+#define LOAD_ELEMENT_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##A = *((__global DATA_TYPE *)(PTR + OFFSET + 10 * STRIDE_Y));
+
+#define LOAD_ELEMENT_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##B = *((__global DATA_TYPE *)(PTR + OFFSET + 11 * STRIDE_Y));
+
+#define LOAD_ELEMENT_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##C = *((__global DATA_TYPE *)(PTR + OFFSET + 12 * STRIDE_Y));
+
+#define LOAD_ELEMENT_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##D = *((__global DATA_TYPE *)(PTR + OFFSET + 13 * STRIDE_Y));
+
+#define LOAD_ELEMENT_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##E = *((__global DATA_TYPE *)(PTR + OFFSET + 14 * STRIDE_Y));
+
+#define LOAD_ELEMENT_16(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##F = *((__global DATA_TYPE *)(PTR + OFFSET + 15 * STRIDE_Y));
+
+/** @}*/ // end of group LOAD_ELEMENT_n
+
+/** Load Scalar as Vector (consecutive elements).
+ * @name LOAD_SCALAR_AS_VECTOR
+ *
+ * Supported cases are M0=1,2,3,...,16 and N0=1,2,3,4,8,16
+ * The data to load is expected to have consecutive names for each row.
+ * E.g., for M0=3, and BASENAME=c, the expected data is c0, c1 and c2.
+ *
+ * @param[in] M0        The number of consecutive rows
+ * @param[in] N0        The number of consecutive columns
+ * @param[in] DATA_TYPE The data type of the target
+ * @param[in] BASENAME  The basename of the result variables
+ * @param[in] PTR       The base pointer for the data
+ * @param[in] OFFSET    The offset within a row
+ * @param[in] STRIDE_Y  The stride in y-axis direction
+ * @{
+ */
+#define LOAD_SCALAR_AS_VECTOR_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) LOAD_ELEMENT_##M0(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)
+#define LOAD_SCALAR_AS_VECTOR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) LOAD_SCALAR_AS_VECTOR_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)
+/** @} */ // end of group LOAD_SCALAR_AS_VECTOR
+
 /** Basic macros to calculate Z offset values from Z0 to Zn-1
  * @name CALCULATE_Z_OFFSET_n
  *
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index 8140b8f2a5..a2a47d7823 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -814,6 +814,301 @@ __kernel void gemmlowp_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 }
+
+#if defined(RESULT_OFFSET) && defined(RESULT_SHIFT) && defined(RESULT_MULTIPLIER)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices with fused output stage using fixed-point arithmetic.
+ *  The LHS matrix is NOT reshaped
+ *  The RHS matrix is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed
+ *
+ * @note The input data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=uchar)
+ * @note The accumulator data type must be passed at compile time using -DACC_DATA_TYPE (i.e. -DACC_DATA_TYPE=uint)
+ * @note The number of columns of LHS matrix must be passed at compile time using -DK (i.e. -DK=64)
+ * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (i.e. -DN0=8, -DK0=4).
+ * @note The number of M0 rows to process must be passed at compile time using -DM0 (i.e. -DM0=2)
+ * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (i.e. -DH0=2)
+ * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
+ * @note Only the following configurations of M0, N0 and K0 are currently supported:
+ *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
+ *  - N0 = 2, 3, 4, 8, 16
+ *  - K0 = 2, 3, 4, 8, 16
+ *  - H0 >= 1
+ *
+ * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
+ *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
+ *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
+ *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
+ *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
+ *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
+ *
+ * @note The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULTIPLIER and -DRESULT_SHIFT
+ * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time
+ * @note The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE
+ * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.
+ *       These values can be used to implement "rectified linear unit" activation functions
+ * @note In case of per-channel quantization of matrix B, -DPER_CHANNEL_QUANTIZATION must be passed at compile time.
+ *
+ * @param[in]  lhs_ptr                                          Pointer to the LHS reshaped matrix. Supported data type: QASYMM8/QASYMM8_SIGNED
+ * @param[in]  lhs_stride_x                                     Stride of the LHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  lhs_step_x                                       src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  lhs_stride_y                                     Stride of the LHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  lhs_step_y                                       src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  lhs_offset_first_element_in_bytes                The offset of the first element in the LHS reshaped matrix
+ * @param[in]  rhs_ptr                                          Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  rhs_stride_x                                     Stride of the RHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  rhs_step_x                                       src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  rhs_stride_y                                     Stride of the RHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  rhs_step_y                                       src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  rhs_offset_first_element_in_bytes                The offset of the first element in the RHS reshaped matrix
+ * @param[out] dst_ptr                                          Pointer to the destination matrix Supported data type: same as @p lhs_ptr
+ * @param[in]  dst_stride_x                                     Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                                       dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                                     Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                                       dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes                The offset of the first element in the destination matrix
+ * @param[in]  lhs_stride_z                                     Stride of the LHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  rhs_stride_z                                     Stride of the RHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  dst_stride_z                                     Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  lhs_cross_plane_pad                              (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
+ * @param[in]  dst_cross_plane_pad                              (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
+ * @param[in]  sum_col_ptr                                      (Optional) Pointer to the source tensor. Supported data type: S32
+ * @param[in]  sum_col_stride_x                                 (Optional) Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  sum_col_step_x                                   (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_col_stride_y                                 (Optional) Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  sum_col_step_y                                   (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  sum_col_offset_first_element_in_bytes            (Optional) The offset of the first element in the source tensor
+ * @param[in]  sum_row_ptr                                      (Optional) Pointer to the source tensor. Supported data type: S32
+ * @param[in]  sum_row_stride_x                                 (Optional) Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  sum_row_step_x                                   (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_row_stride_y                                 (Optional) Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  sum_row_step_y                                   (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  sum_row_offset_first_element_in_bytes            (Optional) The offset of the first element in the source tensor
+ * @param[in]  biases_ptr                                       (Optional) Pointer to the biases tensor. Supported data type: S32
+ * @param[in]  biases_stride_x                                  (Optional) Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                                    (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes             (Optional) The offset of the first element in the biases tensor
+ * @param[in]  result_multipliers_ptr                           (Optional) Pointer to the output multipliers vector for per-channel quantization. Supported data types: S32
+ * @param[in]  result_multipliers_stride_x                      (Optional) Stride of the output multipliers vector in X dimension (in bytes)
+ * @param[in]  result_multipliers_step_x                        (Optional) output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  result_multipliers_offset_first_element_in_bytes (Optional) The offset of the first element in the output multipliers vector
+ * @param[in]  result_shifts_ptr                                (Optional) Pointer to the output shifts vector for per-channel quantization. Supported data types: S32
+ * @param[in]  result_shifts_stride_x                           (Optional) Stride of the output shifts vector in X dimension (in bytes)
+ * @param[in]  result_shifts_step_x                             (Optional) output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  result_shifts_offset_first_element_in_bytes      (Optional) The offset of the first element in the output shifts vector
+ */
+__kernel void gemmlowp_mm_reshaped_only_rhs_t_fused_output_stage_fixedpoint(IMAGE_DECLARATION(lhs),
+                                                                            IMAGE_DECLARATION(rhs),
+                                                                            IMAGE_DECLARATION(dst),
+                                                                            uint lhs_stride_z,
+                                                                            uint rhs_stride_z,
+                                                                            uint dst_stride_z
+#if defined(REINTERPRET_INPUT_AS_3D)
+                                                                            ,
+                                                                            uint lhs_cross_plane_pad
+#endif // REINTERPRET_INPUT_AS_3D
+#if defined(REINTERPRET_OUTPUT_AS_3D)
+                                                                            ,
+                                                                            uint dst_cross_plane_pad
+#endif // REINTERPRET_OUTPUT_AS_3D
+#if defined(A_OFFSET)
+                                                                            ,
+                                                                            IMAGE_DECLARATION(sum_col)
+#endif // defined(A_OFFSET)
+#if defined(B_OFFSET)
+                                                                            ,
+                                                                            IMAGE_DECLARATION(sum_row)
+#endif // defined(B_OFFSET)
+#if defined(ADD_BIAS)
+                                                                            ,
+                                                                            VECTOR_DECLARATION(biases)
+#endif // defined(ADD_BIAS)
+#if defined(PER_CHANNEL_QUANTIZATION)
+                                                                            ,
+                                                                            VECTOR_DECLARATION(result_multipliers),
+                                                                            VECTOR_DECLARATION(result_shifts)
+#endif // defined(PER_CHANNEL_QUANTIZATION)
+                                                                           )
+{
+    // Block size
+#define RHS_BLOCK_SIZE ((K0) * (N0))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (K0)
+#define RHS_STEP_X ((K0) * (H0))
+#define RHS_STEP_LOOP (1)
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X (K0)
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+#if defined(DUMMY_WORK_ITEMS)
+    if((x * N0 >= N) || (y * M0 >= M))
+    {
+        return;
+    }
+#endif // defined(DUMMY_WORK_ITEMS)
+
+    // Compute LHS matrix address
+    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+
+    // Compute RHS matrix address
+    uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X + (x / (uint)H0) * rhs_stride_y;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
+    rhs_offset += z * rhs_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
+    REPEAT_VAR_INIT_TO_CONST(16, uint, zrhs, 0);
+
+#if defined(REINTERPRET_INPUT_AS_3D)
+    // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+
+    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
+    // multiply lhs_stride_z by DEPTH_GEMM3D
+    lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;
+
+#else // defined(REINTERPRET_INPUT_AS_3D)
+
+    // Add offset for batched GEMM
+    lhs_offset += z * lhs_stride_z;
+
+#endif // defined(REINTERPRET_INPUT_AS_3D)
+
+    // Initialize the accumulators
+    REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(ACC_DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(ACC_DATA_TYPE, N0)    c0=0,c1=0,c2=0,... c(N0-1)=0;
+
+    for(int i = 0; i < K; i += K0)
+    {
+        // Load values from LHS matrix
+        LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
+
+        // Load values from RHS matrix
+        LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X, zrhs);
+
+        // Partial matrix multiplication M0,N0,K0
+        ARM_MM_K0XN0XM0(M0, N0, K0, a, b, c);
+
+        lhs_offset += K0;
+        rhs_offset += N0 * RHS_STEP_X * RHS_STEP_LOOP;
+    }
+
+    // Result of MM is of type DATA_TYPE
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(DATA_TYPE) + (y * (uint)M0 * dst_stride_y);
+
+    REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
+
+#if defined(REINTERPRET_OUTPUT_AS_3D)
+    // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+
+    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
+    // multiply dst_stride_z by DEPTH_GEMM3D
+    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
+
+#else // defined(REINTERPRET_OUTPUT_AS_3D)
+
+    // Add offset for batched GEMM
+    dst_addr += z * dst_stride_z;
+
+#endif // defined(REINTERPRET_OUTPUT_AS_3D)
+
+    // Convert result of matrix multiplication to S32
+    REPEAT_VAR_INIT_CONVERT_SAT(M0, VEC_DATA_TYPE(int, N0), c, c_int);
+
+    int batch_id = z;
+#if defined(DEPTH_GEMM3D)
+    batch_id /= (int)DEPTH_GEMM3D;
+#endif // defined(DEPTH_GEMM3D)
+
+    // Offset contribution: c += (A_OFFSET * sum_col) + (B_OFFSET * sum_row) +  K_OFFSET;
+    REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(int, N0), offset_s32_, K_OFFSET);
+
+#if defined(A_OFFSET)
+    // Compute the offset contribution due to A_OFFSET
+    __global uchar *sum_col_addr = sum_col_ptr + sum_col_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(int);
+
+#if defined(SUM_COL_HAS_BATCHES)
+    sum_col_addr += z * sum_col_stride_y;
+#endif // defined(SUM_COL_HAS_BATCHES)
+    VEC_DATA_TYPE(int, N0)
+    a_offset_s32 = VLOAD(N0)(0, (__global int *)sum_col_addr);
+    a_offset_s32 *= (VEC_DATA_TYPE(int, N0))A_OFFSET;
+
+    REPEAT_ADD_VECTOR_TO_VAR(M0, offset_s32_, a_offset_s32);
+#endif // defined(A_OFFSET)
+
+#if defined(B_OFFSET)
+    // Compute the offset contribution due to B_OFFSET
+    __global uchar *sum_row_addr = sum_row_ptr + sum_row_offset_first_element_in_bytes + (y * (uint)M0) * sizeof(int) + z * sum_row_stride_y;
+
+#if defined(HEIGHT_GEMM3D) && defined(DEPTH_GEMM3D)
+    sum_row_addr += (batch_id % (int)DEPTH_GEMM3D) * (int)HEIGHT_GEMM3D * sizeof(int);
+#endif // defined(HEIGHT_GEMM3D) && defined(DEPTH_GEMM3D)
+    LOAD_SCALAR_AS_VECTOR(M0, N0, int, b_offset_s32_, sum_row_addr, 0, sum_row_stride_x);
+
+    REPEAT_MLA_VAR_WITH_CONST_VEC(M0, offset_s32_, b_offset_s32_, (VEC_DATA_TYPE(int, N0))B_OFFSET);
+#endif // defined(B_OFFSET)
+
+#if defined(ADD_BIAS)
+    // Add bias
+    __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(int);
+
+    VEC_DATA_TYPE(int, N0)
+    bias_values = VLOAD(N0)(0, (__global int *)bias_addr);
+    REPEAT_ADD_VECTOR_TO_VAR(M0, offset_s32_, bias_values);
+#endif // defined(ADD_BIAS)
+
+    REPEAT_ADD_TWO_VARS(M0, c_int, offset_s32_);
+
+    // Multiply by result_mult_int and shift
+#if defined(PER_CHANNEL_QUANTIZATION)
+    __global uchar *result_multipliers_addr = result_multipliers_ptr + result_multipliers_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(int);
+    __global uchar *result_shifts_addr      = result_shifts_ptr + result_shifts_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(int);
+
+    VEC_DATA_TYPE(int, N0)
+    res_mul = VLOAD(N0)(0, (__global int *)result_multipliers_addr);
+    VEC_DATA_TYPE(int, N0)
+    res_shift = VLOAD(N0)(0, (__global int *)result_shifts_addr);
+
+    REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL(M0, N0, c_int, res_mul, res_shift);
+#else // defined(PER_CHANNEL_QUANTIZATION)
+
+#if RESULT_SHIFT < 0
+    REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(M0, N0, c_int, RESULT_MULTIPLIER, RESULT_SHIFT);
+#else  // RESULT_SHIFT >= 0
+    REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(M0, N0, c_int, RESULT_MULTIPLIER, RESULT_SHIFT);
+#endif // RESULT_SHIFT < 0
+
+#endif // defined(PER_CHANNEL_QUANTIZATION)
+
+    // Add the offset terms to GEMM's result
+    REPEAT_ADD_CONST_TO_VAR(M0, VEC_DATA_TYPE(int, N0), c_int, RESULT_OFFSET);
+
+#if defined(MIN_BOUND)
+    REPEAT_MAX_CONST_VAR(M0, VEC_DATA_TYPE(int, N0), c_int, MIN_BOUND);
+#endif // defined(MIN_BOUND)
+#if defined(MAX_BOUND)
+    REPEAT_MIN_CONST_VAR(M0, VEC_DATA_TYPE(int, N0), c_int, MAX_BOUND);
+#endif // defined(MAX_BOUND)
+
+    // Convert and store output block (does convert saturate)
+    CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, c_int, dst_addr, dst_stride_y, zout);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+}
+#endif // defined(RESULT_OFFSET) && defined(RESULT_SHIFT) && defined(RESULT_MULTIPLIER)
 #endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(K)
 
 #if defined(M0) && defined(N0) && defined(K0) && defined(K)
@@ -840,7 +1135,7 @@ __kernel void gemmlowp_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
  *
- * @param[in]  lhs_ptr                           Pointer to the LHS reshaped matrix. Supported data type: F16/F32
+ * @param[in]  lhs_ptr                           Pointer to the LHS reshaped matrix. Supported data type: QASYMM8
  * @param[in]  lhs_stride_x                      Stride of the LHS reshaped matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                      Stride of the LHS reshaped matrix in Y dimension (in bytes)
@@ -852,7 +1147,7 @@ __kernel void gemmlowp_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
  * @param[in]  rhs_stride_y                      Stride of the RHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  rhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
- * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: same as @p lhs_ptr
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: S32
  * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
@@ -1389,7 +1684,7 @@ __kernel void gemmlowp_offset_contribution(TENSOR3D_DECLARATION(mm_result)
     vstore4(in_s32, 0, (__global int *)mm_result_addr);
 }
 
-#if defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT)
+#if defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT) && defined(OUTPUT_DATA_TYPE)
 /* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel and it quantizes down to uint8.
  *
  * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), adds to it the offset contribution of matrix A and matrix B and quantizes to uint8 through the output stage.
@@ -1742,7 +2037,7 @@ __kernel void gemmlowp_offset_contribution_quantize_down_fixedpoint(TENSOR3D_DEC
     // Store the result
     vstore4(res, 0, (__global OUTPUT_DATA_TYPE *)dst_addr);
 }
-#endif // defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT)
+#endif // defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT) && defined(OUTPUT_DATA_TYPE)
 
 #endif // defined(K_OFFSET)
 
diff --git a/src/core/CL/cl_kernels/repeat.h b/src/core/CL/cl_kernels/repeat.h
index 691f7aea01..4e761db073 100644
--- a/src/core/CL/cl_kernels/repeat.h
+++ b/src/core/CL/cl_kernels/repeat.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,8 @@
 #ifndef ARM_COMPUTE_REPEAT_H
 #define ARM_COMPUTE_REPEAT_H
 
+#include "helpers.h"
+
 /** Macros that help in loop unrolling */
 //Repeat macros with 3 param, excluding the implicit ID param
 #define REPEAT_3_1(P_X, P_A, P_B, P_C) P_X##_DEF(0, P_A, P_B, P_C)
@@ -76,8 +78,106 @@
 #define REPEAT_DEF_3_N(P_NUM, P_OP, P_A, P_B, P_C) REPEAT_3_##P_NUM(P_OP, P_A, P_B, P_C) //One level of indirection to ensure order of expansion does not affect preprocessing P_NUM
 #define REPEAT_3_N(P_NUM, P_OP, P_A, P_B, P_C) REPEAT_DEF_3_N(P_NUM, P_OP, P_A, P_B, P_C)
 
-//Macro for initializing N variables. generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
+// Repeat macros with 4 param, excluding the implicit ID param
+#define REPEAT_4_1(P_X, P_A, P_B, P_C, P_D) P_X##_DEF(0, P_A, P_B, P_C, P_D)
+#define REPEAT_4_2(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(1, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_1(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_3(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(2, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_2(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_4(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(3, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_3(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_5(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(4, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_4(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_6(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(5, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_5(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_7(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(6, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_6(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_8(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(7, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_7(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_9(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(8, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_8(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_10(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(9, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_9(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_11(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(A, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_10(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_12(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(B, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_11(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_13(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(C, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_12(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_14(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(D, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_13(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_15(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(E, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_14(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_16(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(F, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_15(P_X, P_A, P_B, P_C, P_D)
+
+#define REPEAT_DEF_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D) REPEAT_4_##P_NUM(P_OP, P_A, P_B, P_C, P_D) //One level of indirection to ensure order of expansion does not affect preprocessing P_NUM
+#define REPEAT_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D) REPEAT_DEF_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D)
+
+// Macro for initializing N variables. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
 #define VAR_INIT_TO_CONST_DEF(ID, TYPE, VAR, VAL) TYPE VAR##ID = VAL
 #define REPEAT_VAR_INIT_TO_CONST(N, TYPE, VAR, VAL) REPEAT_3_N(N, VAR_INIT_TO_CONST, TYPE, VAR, VAL)
 
+// Macro for initializing N variables by converting the data type. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
+#define VAR_INIT_CONVERT_SAT_DEF(ID, TYPE_OUT, VAR_IN, VAR_OUT) TYPE_OUT VAR_OUT##ID = CONVERT_SAT(VAR_IN##ID, TYPE_OUT)
+#define REPEAT_VAR_INIT_CONVERT_SAT(N, TYPE_OUT, VAR_IN, VAR_OUT) REPEAT_3_N(N, VAR_INIT_CONVERT_SAT, TYPE_OUT, VAR_IN, VAR_OUT)
+
+// Macro for adding a constant to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ADD_CONST_TO_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID += (TYPE)VAL
+#define REPEAT_ADD_CONST_TO_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, ADD_CONST_TO_VAR, TYPE, VAR, VAL)
+
+// Macro for multiplying N variables (VAR_B) by a constant (VAL) and adding to other N variables (VAR_A). Generates N statements that defines VAR_A##N =RHS_ACCESSOR_DEF(...)
+#define MLA_VAR_WITH_CONST_VEC_DEF(ID, VAR_A, VAR_B, VAL) VAR_A##ID += VAR_B##ID * VAL
+#define REPEAT_MLA_VAR_WITH_CONST_VEC(N, VAR_A, VAR_B, VAL) REPEAT_3_N(N, MLA_VAR_WITH_CONST_VEC, VAR_A, VAR_B, VAL)
+
+// Macro for adding a vector to N-variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ADD_VECTOR_TO_VAR_DEF(ID, TYPE, VAR, VEC) VAR##ID += VEC
+#define REPEAT_ADD_VECTOR_TO_VAR(N, VAR, VEC) REPEAT_3_N(N, ADD_VECTOR_TO_VAR, "", VAR, VEC)
+
+// Macro for adding a two N-variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ADD_TWO_VARS_DEF(ID, TYPE, VAR_A, VAR_B) VAR_A##ID += VAR_B##ID
+#define REPEAT_ADD_TWO_VARS(N, VAR_A, VAR_B) REPEAT_3_N(N, ADD_TWO_VARS, "", VAR_A, VAR_B)
+
+// Macro for performing Max between a constant and N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define MAX_CONST_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID = max(VAR##ID, (TYPE)VAL)
+#define REPEAT_MAX_CONST_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, MAX_CONST_VAR, TYPE, VAR, VAL)
+
+// Macro for performing Min between a constant and N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define MIN_CONST_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID = min(VAR##ID, (TYPE)VAL)
+#define REPEAT_MIN_CONST_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, MIN_CONST_VAR, TYPE, VAR, VAL)
+
+// Macro for performing ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT) VAR##ID = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, SIZE)
+#define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE, SIZE, VAR, RES_MUL, RES_SHIFT)
+
+// Macro for performing ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT) VAR##ID = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, SIZE)
+#define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE, SIZE, VAR, RES_MUL, RES_SHIFT)
+
+// Macro for performing per-channel ASYMM_MULT_BY_QUANT_MULTIPLIER to N variables.
+#define ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT)                     \
+    ({                                                                                                        \
+        VEC_DATA_TYPE(int, N0)                                                                                \
+        VAR##ID_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, N0); \
+        VEC_DATA_TYPE(int, N0)                                                                                \
+        VAR##ID_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, N0);    \
+        VAR##ID           = select(VAR##ID_shift_lt0, VAR##ID_shift_gt0, RES_SHIFT >= 0);                     \
+    })
+#define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL, SIZE, VAR, RES_MUL, RES_SHIFT)
+
 #endif // ARM_COMPUTE_REPEAT_H