COMPMID-1775: Implement CLGEMMReshapeRHSMatrixKernel to reshape the RHS matrix of GEMM/GEMMLowp

Change-Id: I77f2bfcc5d170bcc2428a2f27104942c1ec877d7
Reviewed-on: https://review.mlplatform.org/375
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 754 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 7b98e5ae80..54fc618bdf 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -283,6 +283,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemm_lc_vm_f32", "gemm.cl" },
     { "gemm_transpose1xW", "gemm.cl" },
     { "gemm_reshape_lhs_matrix_nt", "gemm.cl" },
+    { "gemm_reshape_rhs_matrix_nt", "gemm.cl" },
+    { "gemm_reshape_rhs_matrix_t", "gemm.cl" },
     { "gemmlowp_matrix_a_reduction", "gemmlowp.cl" },
     { "gemmlowp_matrix_a_reduction_dot8", "gemmlowp.cl" },
     { "gemmlowp_matrix_b_reduction", "gemmlowp.cl" },
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index cf1e021929..40ee1d45ad 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -252,6 +252,586 @@ __kernel void gemm_reshape_lhs_matrix_nt(TENSOR3D_DECLARATION(src),
 }
 #endif // defined(M0) && defined(K0) && defined(V0) && defined(DATA_TYPE)
 
+#if defined(K0) && defined(N0) && defined(H0) && defined(DATA_TYPE) && defined(SRC_HEIGHT)
+/** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (not transposed) in
+ *  the output matrix unrolling the values.
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=float)
+ * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (i.e. -DSRC_HEIGHT=16)
+ * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (i.e. -DK0=2, -DN0=2).
+ * @note The number of K0xN0 vertical blocks to store on the same output row must be passed at compile time using -DH0 (i.e. -DH0=2)
+ * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time.
+ * @note Only the following values for K0, N0 and H0 are supported:
+ *                                      N0: 2,4,8,16
+ *                                      K0: 1,2,4,8,16
+ *                                      H0: greater than 0
+ *
+ * @param[in]  src_ptr                           Pointer to the source RHS tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  src_stride_x                      Stride of the source RHS tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source RHS tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                      Stride of the source RHS tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source RHS tensor
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_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_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
+ */
+__kernel void gemm_reshape_rhs_matrix_nt(TENSOR3D_DECLARATION(src),
+                                         TENSOR3D_DECLARATION(dst))
+{
+    // Block size
+#define BLOCK_SIZE ((K0) * (N0))
+
+    // Output offset X
+#if defined(INTERLEAVE)
+#define OUTPUT_OFFSET_X (N0)
+#else // defined(INTERLEAVE)
+#define OUTPUT_OFFSET_X (BLOCK_SIZE)
+#endif // defined(INTERLEAVE)
+
+    // Output step X
+#if defined(INTERLEAVE)
+#define OUTPUT_STEP_X (N0) * (H0)
+#else // Do not interleave
+#define OUTPUT_STEP_X (N0)
+#endif // defined(INTERLEAVE)
+
+    // Compute source and destination addresses
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    // ------------------ Compute input/output addresses ---------------------------
+
+    // Compute the input address
+    __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)N0 * sizeof(DATA_TYPE) + y * (uint)K0 * src_stride_y + z * (uint)src_stride_z;
+
+    // Compute the output address
+    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y * (uint)BLOCK_SIZE * (uint)H0 * sizeof(DATA_TYPE)) + ((x % (uint)H0) * (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)) + ((
+                                     x / (uint)H0)
+                                 * (uint)dst_stride_y)
+                                 + z * (uint)dst_stride_z;
+
+    // ---------------------------Load input values --------------------------------
+
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a0 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a1 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a2 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a3 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a4 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a5 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a6 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a7 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a8 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a9 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aA = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aB = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aC = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aD = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aE = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aF = 0;
+
+    // Load values from the RHS matrix
+    a0 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y));
+#if K0 > 1
+    if(y * (uint)K0 + 1 < SRC_HEIGHT)
+    {
+        a1 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y));
+    }
+#endif // K0 > 1
+#if K0 > 2
+    if(y * (uint)K0 + 2 < SRC_HEIGHT)
+    {
+        a2 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y));
+    }
+    if(y * (uint)K0 + 3 < SRC_HEIGHT)
+    {
+        a3 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y));
+    }
+#endif // K0 > 2
+#if K0 > 4
+    if(y * (uint)K0 + 4 < SRC_HEIGHT)
+    {
+        a4 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 4 * src_stride_y));
+    }
+    if(y * (uint)K0 + 5 < SRC_HEIGHT)
+    {
+        a5 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 5 * src_stride_y));
+    }
+    if(y * (uint)K0 + 6 < SRC_HEIGHT)
+    {
+        a6 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 6 * src_stride_y));
+    }
+    if(y * (uint)K0 + 7 < SRC_HEIGHT)
+    {
+        a7 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 7 * src_stride_y));
+    }
+#endif // K0 > 4
+#if K0 > 8
+    if(y * (uint)K0 + 9 < SRC_HEIGHT)
+    {
+        a8 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 8 * src_stride_y));
+    }
+    if(y * (uint)K0 + 9 < SRC_HEIGHT)
+    {
+        a9 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 9 * src_stride_y));
+    }
+    if(y * (uint)K0 + 10 < SRC_HEIGHT)
+    {
+        aA = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 10 * src_stride_y));
+    }
+    if(y * (uint)K0 + 11 < SRC_HEIGHT)
+    {
+        aB = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 11 * src_stride_y));
+    }
+    if(y * (uint)K0 + 12 < SRC_HEIGHT)
+    {
+        aC = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 12 * src_stride_y));
+    }
+    if(y * (uint)K0 + 13 < SRC_HEIGHT)
+    {
+        aD = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 13 * src_stride_y));
+    }
+    if(y * (uint)K0 + 14 < SRC_HEIGHT)
+    {
+        aE = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 14 * src_stride_y));
+    }
+    if(y * (uint)K0 + 15 < SRC_HEIGHT)
+    {
+        aF = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 15 * src_stride_y));
+    }
+#endif // K0 > 8
+
+    // ---------------------------Store output values ------------------------------
+
+    VSTORE(N0)
+    (a0, 0, (__global DATA_TYPE *)(output_ptr + 0 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#if K0 > 1
+    VSTORE(N0)
+    (a1, 0, (__global DATA_TYPE *)(output_ptr + 1 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // K0 > 1
+#if K0 > 2
+    VSTORE(N0)
+    (a2, 0, (__global DATA_TYPE *)(output_ptr + 2 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (a3, 0, (__global DATA_TYPE *)(output_ptr + 3 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // K0 > 2
+#if K0 > 4
+    VSTORE(N0)
+    (a4, 0, (__global DATA_TYPE *)(output_ptr + 4 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (a5, 0, (__global DATA_TYPE *)(output_ptr + 5 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (a6, 0, (__global DATA_TYPE *)(output_ptr + 6 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (a7, 0, (__global DATA_TYPE *)(output_ptr + 7 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 4
+#if K0 > 8
+    VSTORE(N0)
+    (a8, 0, (__global DATA_TYPE *)(output_ptr + 8 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (a9, 0, (__global DATA_TYPE *)(output_ptr + 9 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (aA, 0, (__global DATA_TYPE *)(output_ptr + 10 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (aB, 0, (__global DATA_TYPE *)(output_ptr + 11 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (aC, 0, (__global DATA_TYPE *)(output_ptr + 12 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (aD, 0, (__global DATA_TYPE *)(output_ptr + 13 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (aE, 0, (__global DATA_TYPE *)(output_ptr + 14 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(N0)
+    (aF, 0, (__global DATA_TYPE *)(output_ptr + 15 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 8
+
+#undef BLOCK_SIZE
+#undef OUTPUT_OFFSET_X
+#undef OUTPUT_STEP_X
+}
+
+#if defined(TRANSPOSE)
+/** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (transposed) in
+ *  the output matrix unrolling the values.
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=float)
+ * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (i.e. -DSRC_HEIGHT=16)
+ * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (i.e. -DK0=2, -DN0=2).
+ * @note The number of K0xN0 vertical blocks to store on the same output row must be passed at compile time using -DH0 (i.e. -DH0=2)
+ * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time.
+ * @note The option -DTRANSPOSE must passed at compile time.
+ * @note Only the following values for K0, N0 and H0 are supported:
+ *                                      N0: 2,4,8,16
+ *                                      K0: 4,8,16
+ *                                      H0: greater than 0
+ *
+ * @param[in]  src_ptr                           Pointer to the source RHS tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  src_stride_x                      Stride of the source RHS tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source RHS tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                      Stride of the source RHS tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source RHS tensor
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_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_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
+ */
+__kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src),
+                                        TENSOR3D_DECLARATION(dst))
+{
+    // Block size
+#define BLOCK_SIZE ((K0) * (N0))
+
+    // Output offset X
+#if defined(INTERLEAVE)
+#define OUTPUT_OFFSET_X (K0)
+#else // defined(INTERLEAVE)
+#define OUTPUT_OFFSET_X (BLOCK_SIZE)
+#endif // defined(INTERLEAVE)
+
+    // Output step X
+#if defined(INTERLEAVE)
+#define OUTPUT_STEP_X (K0) * (H0)
+#else // Do not interleave
+#define OUTPUT_STEP_X (K0)
+#endif // defined(INTERLEAVE)
+
+    // Compute source and destination addresses
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    // ------------------ Compute input/output addresses ---------------------------
+
+    // Compute the input address
+    __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)N0 * sizeof(DATA_TYPE) + y * (uint)K0 * src_stride_y + z * (uint)src_stride_z;
+
+    // Compute the output address
+    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y * (uint)BLOCK_SIZE * (uint)H0 * sizeof(DATA_TYPE)) + ((x % H0) * (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)) + ((x /
+                                 (uint)H0) * (uint)dst_stride_y) + z * (uint)dst_stride_z;
+
+    // ---------------------------Load input values --------------------------------
+
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a0 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a1 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a2 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a3 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a4 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a5 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a6 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a7 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a8 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    a9 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aA = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aB = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aC = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aD = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aE = 0;
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    aF = 0;
+
+    // Load values from the RHS matrix
+    a0 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y));
+    if(y * (uint)K0 + 1 < SRC_HEIGHT)
+    {
+        a1 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y));
+    }
+    if(y * (uint)K0 + 2 < SRC_HEIGHT)
+    {
+        a2 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y));
+    }
+    if(y * (uint)K0 + 3 < SRC_HEIGHT)
+    {
+        a3 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y));
+    }
+#if K0 > 4
+    if(y * (uint)K0 + 4 < SRC_HEIGHT)
+    {
+        a4 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 4 * src_stride_y));
+    }
+    if(y * (uint)K0 + 5 < SRC_HEIGHT)
+    {
+        a5 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 5 * src_stride_y));
+    }
+    if(y * (uint)K0 + 6 < SRC_HEIGHT)
+    {
+        a6 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 6 * src_stride_y));
+    }
+    if(y * (uint)K0 + 7 < SRC_HEIGHT)
+    {
+        a7 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 7 * src_stride_y));
+    }
+#endif // K0 > 4
+#if K0 > 8
+    if(y * (uint)K0 + 9 < SRC_HEIGHT)
+    {
+        a8 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 8 * src_stride_y));
+    }
+    if(y * (uint)K0 + 9 < SRC_HEIGHT)
+    {
+        a9 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 9 * src_stride_y));
+    }
+    if(y * (uint)K0 + 10 < SRC_HEIGHT)
+    {
+        aA = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 10 * src_stride_y));
+    }
+    if(y * (uint)K0 + 11 < SRC_HEIGHT)
+    {
+        aB = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 11 * src_stride_y));
+    }
+    if(y * (uint)K0 + 12 < SRC_HEIGHT)
+    {
+        aC = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 12 * src_stride_y));
+    }
+    if(y * (uint)K0 + 13 < SRC_HEIGHT)
+    {
+        aD = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 13 * src_stride_y));
+    }
+    if(y * (uint)K0 + 14 < SRC_HEIGHT)
+    {
+        aE = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 14 * src_stride_y));
+    }
+    if(y * (uint)K0 + 15 < SRC_HEIGHT)
+    {
+        aF = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 15 * src_stride_y));
+    }
+#endif // K0 > 8
+
+    // ---------------------------Transpose the block ------------------------------
+
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res0 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res1 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res2 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res3 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res4 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res5 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res6 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res7 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res8 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    res9 = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    resA = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    resB = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    resC = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    resD = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    resE = 0;
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    resF = 0;
+
+#if K0 == 4
+    // This part computes the following transpositions:
+    // 4x2 -> 2x4
+    // 4x4 -> 4x4
+    // 4x8 -> 8x4
+    // 4x16 -> 16x4
+    res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0);
+    res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1);
+#if N0 > 2
+    res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2);
+    res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3);
+#endif // N0 > 2
+#if N0 > 4
+    res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4);
+    res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5);
+    res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6);
+    res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7);
+#endif // N0 > 4
+#if N0 > 8
+    res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8);
+    res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9);
+    resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA);
+    resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB);
+    resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC);
+    resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD);
+    resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE);
+    resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF);
+#endif // N0 > 8
+
+#elif K0 == 8 // N0 == 3
+    // This part computes the following transpositions:
+    // 8x2 -> 2x8
+    // 8x4 -> 4x8
+    // 8x8 -> 8x8
+    // 8x16 -> 16x8
+    res0                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0, a4.s0, a5.s0, a6.s0, a7.s0);
+    res1                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1, a4.s1, a5.s1, a6.s1, a7.s1);
+#if N0 > 2
+    res2                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2, a4.s2, a5.s2, a6.s2, a7.s2);
+    res3                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3, a4.s3, a5.s3, a6.s3, a7.s3);
+#endif // N0 > 2
+#if N0 > 4
+    res4                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4, a4.s4, a5.s4, a6.s4, a7.s4);
+    res5                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5, a4.s5, a5.s5, a6.s5, a7.s5);
+    res6                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6, a4.s6, a5.s6, a6.s6, a7.s6);
+    res7                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7, a4.s7, a5.s7, a6.s7, a7.s7);
+#endif // N0 > 4
+#if N0 > 8
+    res8                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8, a4.s8, a5.s8, a6.s8, a7.s8);
+    res9                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9, a4.s9, a5.s9, a6.s9, a7.s9);
+    resA                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA, a4.sA, a5.sA, a6.sA, a7.sA);
+    resB                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB, a4.sB, a5.sB, a6.sB, a7.sB);
+    resC                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC, a4.sC, a5.sC, a6.sC, a7.sC);
+    resD                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD, a4.sD, a5.sD, a6.sD, a7.sD);
+    resE                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE, a4.sE, a5.sE, a6.sE, a7.sE);
+    resF                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF, a4.sF, a5.sF, a6.sF, a7.sF);
+#endif // N0 > 8
+
+#elif K0 == 16 // N0 == 16
+
+    // This part computes the following transpositions:
+    // 16x2 -> 2x16
+    // 16x4 -> 4x16
+    // 16x8 -> 8x16
+    // 16x16 -> 16x16
+    res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0, a4.s0, a5.s0, a6.s0, a7.s0,
+                                          a8.s0, a9.s0, aA.s0, aB.s0, aC.s0, aD.s0, aE.s0, aF.s0);
+    res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1, a4.s1, a5.s1, a6.s1, a7.s1,
+                                          a8.s1, a9.s1, aA.s1, aB.s1, aC.s1, aD.s1, aE.s1, aF.s1);
+#if N0 > 2
+    res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2, a4.s2, a5.s2, a6.s2, a7.s2,
+                                          a8.s2, a9.s2, aA.s2, aB.s2, aC.s2, aD.s2, aE.s2, aF.s2);
+    res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3, a4.s3, a5.s3, a6.s3, a7.s3,
+                                          a8.s3, a9.s3, aA.s3, aB.s3, aC.s3, aD.s3, aE.s3, aF.s3);
+#endif // N0 > 2
+#if N0 > 4
+    res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4, a4.s4, a5.s4, a6.s4, a7.s4,
+                                          a8.s4, a9.s4, aA.s4, aB.s4, aC.s4, aD.s4, aE.s4, aF.s4);
+    res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5, a4.s5, a5.s5, a6.s5, a7.s5,
+                                          a8.s5, a9.s5, aA.s5, aB.s5, aC.s5, aD.s5, aE.s5, aF.s5);
+    res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6, a4.s6, a5.s6, a6.s6, a7.s6,
+                                          a8.s6, a9.s6, aA.s6, aB.s6, aC.s6, aD.s6, aE.s6, aF.s6);
+    res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7, a4.s7, a5.s7, a6.s7, a7.s7,
+                                          a8.s7, a9.s7, aA.s7, aB.s7, aC.s7, aD.s7, aE.s7, aF.s7);
+#endif // N0 > 4
+#if N0 > 8
+    res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8, a4.s8, a5.s8, a6.s8, a7.s8,
+                                          a8.s8, a9.s8, aA.s8, aB.s8, aC.s8, aD.s8, aE.s8, aF.s8);
+    res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9, a4.s9, a5.s9, a6.s9, a7.s9,
+                                          a8.s9, a9.s9, aA.s9, aB.s9, aC.s9, aD.s9, aE.s9, aF.s9);
+    resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA, a4.sA, a5.sA, a6.sA, a7.sA,
+                                          a8.sA, a9.sA, aA.sA, aB.sA, aC.sA, aD.sA, aE.sA, aF.sA);
+    resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB, a4.sB, a5.sB, a6.sB, a7.sB,
+                                          a8.sB, a9.sB, aA.sB, aB.sB, aC.sB, aD.sB, aE.sB, aF.sB);
+    resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC, a4.sC, a5.sC, a6.sC, a7.sC,
+                                          a8.sC, a9.sC, aA.sC, aB.sC, aC.sC, aD.sC, aE.sC, aF.sC);
+    resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD, a4.sD, a5.sD, a6.sD, a7.sD,
+                                          a8.sD, a9.sD, aA.sD, aB.sD, aC.sD, aD.sD, aE.sD, aF.sD);
+    resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE, a4.sE, a5.sE, a6.sE, a7.sE,
+                                          a8.sE, a9.sE, aA.sE, aB.sE, aC.sE, aD.sE, aE.sE, aF.sE);
+    resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF, a4.sF, a5.sF, a6.sF, a7.sF,
+                                          a8.sF, a9.sF, aA.sF, aB.sF, aC.sF, aD.sF, aE.sF, aF.sF);
+#endif // N0 > 8
+
+#else // N0 == 16
+#error "Not supported N0 value"
+#endif // N0 > 2
+
+    // ---------------------------Store the output values ------------------------------
+
+    VSTORE(K0)
+    (res0, 0, (__global DATA_TYPE *)(output_ptr + 0 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (res1, 0, (__global DATA_TYPE *)(output_ptr + 1 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#if N0 > 2
+    VSTORE(K0)
+    (res2, 0, (__global DATA_TYPE *)(output_ptr + 2 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (res3, 0, (__global DATA_TYPE *)(output_ptr + 3 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 2
+#if N0 > 4
+    VSTORE(K0)
+    (res4, 0, (__global DATA_TYPE *)(output_ptr + 4 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (res5, 0, (__global DATA_TYPE *)(output_ptr + 5 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (res6, 0, (__global DATA_TYPE *)(output_ptr + 6 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (res7, 0, (__global DATA_TYPE *)(output_ptr + 7 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 4
+#if N0 > 8
+    VSTORE(K0)
+    (res8, 0, (__global DATA_TYPE *)(output_ptr + 8 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (res9, 0, (__global DATA_TYPE *)(output_ptr + 9 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (resA, 0, (__global DATA_TYPE *)(output_ptr + 10 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (resB, 0, (__global DATA_TYPE *)(output_ptr + 11 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (resC, 0, (__global DATA_TYPE *)(output_ptr + 12 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (resD, 0, (__global DATA_TYPE *)(output_ptr + 13 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (resE, 0, (__global DATA_TYPE *)(output_ptr + 14 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+    VSTORE(K0)
+    (resF, 0, (__global DATA_TYPE *)(output_ptr + 15 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 8
+
+#undef BLOCK_SIZE
+#undef OUTPUT_OFFSET_X
+#undef OUTPUT_STEP_X
+}
+#endif // defined(TRANSPOSE)
+#endif // defined(K0) && defined(N0) && defined(H0) && defined(DATA_TYPE) && defined(SRC_HEIGHT)
+
 #if defined(TRANSPOSE_W) && defined(MULT_TRANSPOSE1XW_WIDTH)
 
 #if ELEMENT_SIZE == 1
diff --git a/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp b/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp
new file mode 100644
index 0000000000..7cfa234b2e
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp
@@ -0,0 +1,172 @@
+/*
+ * Copyright (c) 2018 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/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const GEMMRHSMatrixInfo &rhs_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.h0 == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.n0 & (rhs_info.n0 - 1)), "Only power of two values are allowed for n0");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.k0 & (rhs_info.k0 - 1)), "Only power of two values are allowed for k0");
+    ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 > 16);
+    ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16);
+    ARM_COMPUTE_RETURN_ERROR_ON((rhs_info.k0 == 1) && (rhs_info.transpose));
+    ARM_COMPUTE_RETURN_ERROR_ON((rhs_info.k0 == 2) && (rhs_info.transpose));
+
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::U8, DataType::S8,
+                                                         DataType::U16, DataType::S16, DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_rhs_reshaped_shape(*input, rhs_info));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const GEMMRHSMatrixInfo &rhs_info)
+{
+    const unsigned int num_elems_processed_per_iteration_x = rhs_info.n0;
+    const unsigned int num_elems_processed_per_iteration_y = rhs_info.k0;
+    bool               window_changed                      = false;
+
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*input, rhs_info)));
+
+    // Configure window
+    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+    AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+    AccessWindowStatic    output_access(output, 0, 0, output->dimension(0), output->dimension(1));
+
+    window_changed = update_window_and_padding(win, input_access);
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
+
+    // Collapse along the Z direction
+    // This collapse needs to be here in order to tune the Z dimension of LWS
+    Window collapsed = win.collapse(win, Window::DimZ);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, collapsed);
+}
+} // namespace
+
+CLGEMMReshapeRHSMatrixKernel::CLGEMMReshapeRHSMatrixKernel()
+    : _input(nullptr), _output(nullptr)
+{
+}
+
+void CLGEMMReshapeRHSMatrixKernel::configure(const ICLTensor *input, ICLTensor *output, const GEMMRHSMatrixInfo &rhs_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), rhs_info));
+
+    _input  = input;
+    _output = output;
+
+    // Create build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
+    build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
+    build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
+    build_opts.add_option_if(rhs_info.transpose, "-DTRANSPOSE");
+    build_opts.add_option_if(rhs_info.interleave, "-DINTERLEAVE");
+    build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
+
+    switch(input->info()->element_size())
+    {
+        case 1:
+            build_opts.add_option("-DDATA_TYPE=uchar");
+            break;
+        case 2:
+            build_opts.add_option("-DDATA_TYPE=ushort");
+            break;
+        case 4:
+            build_opts.add_option("-DDATA_TYPE=uint");
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Data type not supported");
+    }
+
+    std::string kernel_name("gemm_reshape_rhs_matrix_");
+    kernel_name += rhs_info.transpose ? "t" : "nt";
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info(), rhs_info);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+}
+
+Status CLGEMMReshapeRHSMatrixKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const GEMMRHSMatrixInfo &rhs_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, rhs_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), rhs_info).first);
+
+    return Status{};
+}
+
+void CLGEMMReshapeRHSMatrixKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window slice = window.first_slice_window_3D();
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice);
+        enqueue(queue, *this, slice);
+    }
+    while(window.slide_window_slice_3D(slice));
+}
\ No newline at end of file