COMPMID-2000: Implement CLGEMMMatrixMultiplyReshapedOnlyRHS - Transposed

Change-Id: I364c7ec5a43ad391a73429489802b0e679ee0c6e
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/732
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

8 files changed, 1606 insertions, 7 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index cc4888c663..e0d124bff9 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -81,6 +81,7 @@
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixAdditionKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h
new file mode 100644
index 0000000000..74715949f2
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h
@@ -0,0 +1,89 @@
+/*
+ * Copyright (c) 2019 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_CLGEMMMATRIXMULTIPLYRESHAPEDONLYRHSKERNEL_H__
+#define __ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDONLYRHSKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** OpenCL kernel to multiply matrices when only the input matrix RHS (input1) has been reshaped
+ *
+ * @note The input matrix input1 must be reshaped through @ref CLGEMMReshapeRHSMatrixKernel
+ */
+class CLGEMMMatrixMultiplyReshapedOnlyRHSKernel : public ICLKernel
+{
+public:
+    /** Default Constructor */
+    CLGEMMMatrixMultiplyReshapedOnlyRHSKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMMatrixMultiplyReshapedOnlyRHSKernel(const CLGEMMMatrixMultiplyReshapedOnlyRHSKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMMatrixMultiplyReshapedOnlyRHSKernel &operator=(const CLGEMMMatrixMultiplyReshapedOnlyRHSKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLGEMMMatrixMultiplyReshapedOnlyRHSKernel(CLGEMMMatrixMultiplyReshapedOnlyRHSKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLGEMMMatrixMultiplyReshapedOnlyRHSKernel &operator=(CLGEMMMatrixMultiplyReshapedOnlyRHSKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  input0    Input tensor containing the LHS matrix. Data type supported: F32/F16
+     * @param[in]  input1    Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0
+     * @param[out] output    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
+     * @param[in]  alpha     Weight of the matrix product
+     * @param[in]  lhs_info  LHS matrix information used to retrieve the number of rows to be processed by each thread
+     * @param[in]  rhs_info  RHS matrix information used for reshaping the input1 tensor
+     * @param[in]  gemm_info GEMM information used to retrieve the original dimensions of the input matrices
+     */
+    void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                   const GEMMReshapeInfo &gemm_info);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
+     *
+     * @param[in] input0    Input tensor info for the LHS matrix. Data type supported: F32/F16
+     * @param[in] input1    Input tensor info for the RHS reshaped matrix. Data type supported: same as @p input0
+     * @param[in] output    Output tensor info. Data type supported: same as @p input0
+     * @param[in] alpha     Weight of the matrix product
+     * @param[in] lhs_info  LHS matrix information used to retrieve the number of rows to be processed by each thread
+     * @param[in] rhs_info  RHS matrix information used for reshaping the input1 tensor
+     * @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                           const GEMMReshapeInfo &gemm_info);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor *_input0;
+    const ICLTensor *_input1;
+    ICLTensor       *_output;
+    bool             _slide_matrix_b;
+    bool             _reinterpret_input_as_3d;
+    bool             _reinterpret_output_as_3d;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDONLYRHSKERNEL_H__*/
\ No newline at end of file
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 4ecb885440..a86d8d0414 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -284,6 +284,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemm_mm_floating_point_f32_bifrost", "gemm.cl" },
     { "gemm_mm_floating_point_f32_bifrost_1000", "gemm.cl" },
     { "gemm_mm_reshaped_lhs_nt_rhs_t", "gemm.cl" },
+    { "gemm_mm_reshaped_only_rhs_t", "gemm.cl" },
     { "gemm_lc_vm_f32", "gemm.cl" },
     { "gemm_transpose1xW", "gemm.cl" },
     { "gemm_reshape_lhs_matrix_nt", "gemm.cl" },
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index 4736f80d9b..a245aeead3 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -1128,6 +1128,634 @@ __kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src),
 #endif // defined(TRANSPOSE)
 #endif // defined(K0) && defined(N0) && defined(H0) && defined(DATA_TYPE) && defined(SRC_HEIGHT)
 
+#if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(K)
+
+#define CONCAT(a, b) a##b
+
+#define ARM_DOT1(a, b, c) \
+    ({                    \
+        c = fma(a, b, c); \
+    })
+#define ARM_DOT2(a, b, c)       \
+    ({                          \
+        c = fma(a.s0, b.s0, c); \
+        c = fma(a.s1, b.s1, c); \
+    })
+#define ARM_DOT3(a, b, c)           \
+    ({                              \
+        ARM_DOT2(a, b, c);          \
+        c = fma((a.s2), (b.s2), c); \
+    })
+#define ARM_DOT4(a, b, c)           \
+    ({                              \
+        ARM_DOT3(a, b, c);          \
+        c = fma((a.s3), (b.s3), c); \
+    })
+#define ARM_DOT8(a, b, c)            \
+    ({                               \
+        ARM_DOT4((a.lo), (b.lo), c); \
+        ARM_DOT4((a.hi), (b.hi), c); \
+    })
+#define ARM_DOT16(a, b, c)           \
+    ({                               \
+        ARM_DOT8((a.lo), (b.lo), c); \
+        ARM_DOT8((a.hi), (b.hi), c); \
+    })
+
+#if N0 == 2
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+    })
+#elif N0 == 3 // N0 == 3
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+    })
+#elif N0 == 4 // N0 == 4
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##3), (c.s3));     \
+    })
+#elif N0 == 8 // N0 == 8
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##3), (c.s3));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##4), (c.s4));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##5), (c.s5));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##6), (c.s6));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##7), (c.s7));     \
+    })
+#elif N0 == 16 // N0 == 16
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##3), (c.s3));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##4), (c.s4));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##5), (c.s5));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##6), (c.s6));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##7), (c.s7));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##8), (c.s8));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##9), (c.s9));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##A), (c.sA));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##B), (c.sB));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##C), (c.sC));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##D), (c.sD));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##E), (c.sE));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##F), (c.sF));     \
+    })
+#else // N0 not supported
+#error "N0 value not supported"
+#endif // N0 conditions
+
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
+ *  The LHS matrix is NOT reshaped
+ *  The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed
+ *
+ * @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
+ *
+ * @param[in]  lhs_ptr                           Pointer to the LHS reshaped matrix. Supported data type: F16/F32
+ * @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)
+ */
+__kernel void gemm_mm_reshaped_only_rhs_t(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
+                                         )
+{
+// 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);
+
+    // 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 * sizeof(DATA_TYPE) + (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, zin, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
+
+#if defined(REINTERPRET_INPUT_AS_3D)
+    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
+    // in order to take into account the presence of possible cross plane paddings
+    //
+    //  |                  |
+    //  |      plane0      |
+    //  |                  |
+    //  |__________________|
+    //  |******************|
+    //  |  cross_plane_pad |
+    //  |******************|
+    //  |                  |
+    //  |      plane1      |
+    //  |                  |
+    //  |__________________|
+
+    // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    zin0 = (0 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zin0 = min((uint)(DEPTH_GEMM3D - 1), zin0);
+    zin0 *= (lhs_cross_plane_pad * lhs_stride_y);
+#if M0 > 1
+    zin1 = (1 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zin1 = min((uint)(DEPTH_GEMM3D - 1), zin1);
+    zin1 *= (lhs_cross_plane_pad * lhs_stride_y);
+#endif // M0 > 1
+#if M0 > 2
+    zin2 = (2 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zin2 = min((uint)(DEPTH_GEMM3D - 1), zin2);
+    zin2 *= (lhs_cross_plane_pad * lhs_stride_y);
+#endif // M0 > 2
+#if M0 > 3
+    zin3 = (3 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zin3 = min((uint)(DEPTH_GEMM3D - 1), zin3);
+    zin3 *= (lhs_cross_plane_pad * lhs_stride_y);
+#endif // M0 > 3
+#if M0 > 4
+    zin4 = (4 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zin4 = min((uint)(DEPTH_GEMM3D - 1), zin4);
+    zin4 *= (lhs_cross_plane_pad * lhs_stride_y);
+#endif // M0 > 4
+#if M0 > 5
+    zin5 = (5 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zin5 = min((uint)(DEPTH_GEMM3D - 1), zin5);
+    zin5 *= (lhs_cross_plane_pad * lhs_stride_y);
+#endif // M0 > 5
+#if M0 > 6
+    zin6 = (6 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zin6 = min((uint)(DEPTH_GEMM3D - 1), zin6);
+    zin6 *= (lhs_cross_plane_pad * lhs_stride_y);
+#endif // M0 > 6
+#if M0 > 7
+    zin7 = (7 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zin7 = min((uint)(DEPTH_GEMM3D - 1), zout7);
+    zin7 *= (lhs_cross_plane_pad * lhs_stride_y);
+#endif // M0 > 7
+
+    // 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(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0)    c0=0,c1=0,c2=0,... c(M0-1)=0;
+
+    int i = 0;
+    for(; i <= (K - K0); i += K0)
+    {
+        // Supported cases (M0, K0):
+        // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
+        // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
+        // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
+        // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
+        // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
+        // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
+        // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
+        // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
+        // Load values from LHS matrix
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a0 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0));
+#if M0 > 1
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a1 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
+#endif // M0 > 1
+#if M0 > 2
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a2 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
+#endif // M0 > 2
+#if M0 > 3
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a3 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
+#endif // M0 > 3
+#if M0 > 4
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a4 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
+#endif // M0 > 4
+#if M0 > 5
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a5 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
+#endif // M0 > 5
+#if M0 > 6
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a6 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
+#endif // M0 > 6
+#if M0 > 7
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a7 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
+#endif // M0 > 7
+
+        // Load values from RHS matrix
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b0 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b1 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 1 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#if N0 > 2
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b2 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 2 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 2
+#if N0 > 3
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b3 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 3 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 3
+#if N0 > 4
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b4 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 4 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b5 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 5 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b6 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 6 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b7 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 7 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 4
+#if N0 > 8
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b8 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 8 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b9 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 9 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bA = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 10 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bB = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 11 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bC = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 12 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bD = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 13 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bE = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 14 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bF = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 15 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 8
+
+        // Accumulate
+        ARM_DOT_K0XN0(K0, a0, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(K0, a1, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(K0, a2, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(K0, a3, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(K0, a4, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(K0, a5, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(K0, a6, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(K0, a7, b, c7);
+#endif // M0 > 7
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        rhs_offset += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE);
+    }
+
+    // Left-over accumulations
+    for(; i < K; ++i)
+    {
+        // Supported cases (M0, K0):
+        // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
+        // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
+        // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
+        // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
+        // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
+        // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
+        // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
+        // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
+        // Load values from LHS matrix
+        DATA_TYPE a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0));
+#if M0 > 1
+        DATA_TYPE a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
+#endif // M0 > 1
+#if M0 > 2
+        DATA_TYPE a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
+#endif // M0 > 2
+#if M0 > 3
+        DATA_TYPE a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
+#endif // M0 > 3
+#if M0 > 4
+        DATA_TYPE a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
+#endif // M0 > 4
+#if M0 > 5
+        DATA_TYPE a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
+#endif // M0 > 5
+#if M0 > 6
+        DATA_TYPE a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
+#endif // M0 > 6
+#if M0 > 7
+        DATA_TYPE a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
+#endif // M0 > 7
+
+        // Load values from RHS matrix
+        DATA_TYPE b0 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE b1 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 1 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#if N0 > 2
+        DATA_TYPE b2 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 2 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 2
+#if N0 > 3
+        DATA_TYPE b3 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 3 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 3
+#if N0 > 4
+        DATA_TYPE b4 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 4 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE b5 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 5 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE b6 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 6 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE b7 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 7 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 4
+#if N0 > 8
+        DATA_TYPE b8 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 8 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE b9 = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 9 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE bA = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 10 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE bB = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 11 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE bC = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 12 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE bD = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 13 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE bE = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 14 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        DATA_TYPE bF = *((__global DATA_TYPE *)(rhs_ptr + rhs_offset + 15 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 8
+
+        // Accumulate
+        ARM_DOT_K0XN0(1, a0, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(1, a1, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(1, a2, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(1, a3, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(1, a4, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(1, a5, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(1, a6, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(1, a7, b, c7);
+#endif // M0 > 7
+
+        lhs_offset += sizeof(DATA_TYPE);
+        rhs_offset += sizeof(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)
+    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
+    // in order to take into account the presence of possible cross plane paddings
+    //
+    //  |                  |
+    //  |      plane0      |
+    //  |                  |
+    //  |__________________|
+    //  |******************|
+    //  |  cross_plane_pad |
+    //  |******************|
+    //  |                  |
+    //  |      plane1      |
+    //  |                  |
+    //  |__________________|
+
+    // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    zout0 = (0 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout0 = min((uint)(DEPTH_GEMM3D - 1), zout0);
+    zout0 *= (dst_cross_plane_pad * dst_stride_y);
+#if M0 > 1
+    zout1 = (1 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout1 = min((uint)(DEPTH_GEMM3D - 1), zout1);
+    zout1 *= (dst_cross_plane_pad * dst_stride_y);
+#endif // M0 > 1
+#if M0 > 2
+    zout2 = (2 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout2 = min((uint)(DEPTH_GEMM3D - 1), zout2);
+    zout2 *= (dst_cross_plane_pad * dst_stride_y);
+#endif // M0 > 2
+#if M0 > 3
+    zout3 = (3 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout3 = min((uint)(DEPTH_GEMM3D - 1), zout3);
+    zout3 *= (dst_cross_plane_pad * dst_stride_y);
+#endif // M0 > 3
+#if M0 > 4
+    zout4 = (4 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout4 = min((uint)(DEPTH_GEMM3D - 1), zout4);
+    zout4 *= (dst_cross_plane_pad * dst_stride_y);
+#endif // M0 > 4
+#if M0 > 5
+    zout5 = (5 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout5 = min((uint)(DEPTH_GEMM3D - 1), zout5);
+    zout5 *= (dst_cross_plane_pad * dst_stride_y);
+#endif // M0 > 5
+#if M0 > 6
+    zout6 = (6 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout6 = min((uint)(DEPTH_GEMM3D - 1), zout6);
+    zout6 *= (dst_cross_plane_pad * dst_stride_y);
+#endif // M0 > 6
+#if M0 > 7
+    zout7 = (7 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout7 = min((uint)(DEPTH_GEMM3D - 1), zout7);
+    zout7 *= (dst_cross_plane_pad * dst_stride_y);
+#endif // M0 > 7
+
+    // 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)
+
+    // Multiply by the weight of matrix-matrix product and store the result
+#if defined(ALPHA)
+    c0 = c0 * (DATA_TYPE)ALPHA;
+#if M0 > 1
+    c1 = c1 * (DATA_TYPE)ALPHA;
+#endif // M0 > 1
+#if M0 > 2
+    c2 = c2 * (DATA_TYPE)ALPHA;
+#endif // M0 > 2
+#if M0 > 3
+    c3 = c3 * (DATA_TYPE)ALPHA;
+#endif // M0 > 3
+#if M0 > 4
+    c4 = c4 * (DATA_TYPE)ALPHA;
+#endif // M0 > 4
+#if M0 > 5
+    c5 = c5 * (DATA_TYPE)ALPHA;
+#endif // M0 > 5
+#if M0 > 6
+    c6 = c6 * (DATA_TYPE)ALPHA;
+#endif // M0 > 5
+#if M0 > 7
+    c7 = c7 * (DATA_TYPE)ALPHA;
+#endif // M0 > 7
+#endif // defined(ALPHA)
+
+    // Store output block
+    VSTORE(N0)
+    (c0, 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y + zout0));
+#if M0 > 1
+    VSTORE(N0)
+    (c1, 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y + zout1));
+#endif // M0 > 1
+#if M0 > 2
+    VSTORE(N0)
+    (c2, 0, (__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y + zout2));
+#endif // M0 > 2
+#if M0 > 3
+    VSTORE(N0)
+    (c3, 0, (__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y + zout3));
+#endif // M0 > 3
+#if M0 > 4
+    VSTORE(N0)
+    (c4, 0, (__global DATA_TYPE *)(dst_addr + 4 * dst_stride_y + zout4));
+#endif // M0 > 4
+#if M0 > 5
+    VSTORE(N0)
+    (c5, 0, (__global DATA_TYPE *)(dst_addr + 5 * dst_stride_y + zout5));
+#endif // M0 > 5
+#if M0 > 6
+    VSTORE(N0)
+    (c6, 0, (__global DATA_TYPE *)(dst_addr + 6 * dst_stride_y + zout6));
+#endif // M0 > 6
+#if M0 > 7
+    VSTORE(N0)
+    (c7, 0, (__global DATA_TYPE *)(dst_addr + 7 * dst_stride_y + zout7));
+#endif // M0 > 7
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+}
+#endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(K)
+
 #if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE)
 
 #if K0 == 2
@@ -1254,7 +1882,7 @@ __kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src),
  * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time.
  * @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 = 2, 3, 4, 5, 6, 7, 8
+ *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
  *  - N0 = 2, 3, 4, 8, 16
  *  - K0 = 2, 3, 4, 8, 16
  *
@@ -1348,11 +1976,14 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs),
     for(int i = 0; i < k; i += K0)
     {
         // Supported cases (M0, K0):
-        // 2,4 - 2,8 - 2,16
-        // 3,4 - 3,8 - 3,16
-        // 4,4 - 4,8 - 4,16
-        // 5,4 - 5,8 - 5,16
-        // 6,4 - 6,8 - 6,16
+        // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
+        // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
+        // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
+        // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
+        // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
+        // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
+        // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
+        // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
         // Load values from LHS matrix
         VEC_DATA_TYPE(DATA_TYPE, K0)
         a0 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_addr + 0 * LHS_STEP_X * sizeof(DATA_TYPE)));
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index 56e9db5816..beff7ae8c4 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -33,7 +33,6 @@
 #include "arm_compute/core/Types.h"
 #include "support/ToolchainSupport.h"
 
-#include "arm_compute/runtime/CL/CLScheduler.h"
 #include <tuple>
 
 using namespace arm_compute;
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
new file mode 100644
index 0000000000..6dc2a9e610
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
@@ -0,0 +1,315 @@
+/*
+ * Copyright (c) 2019 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/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.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/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/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "support/ToolchainSupport.h"
+
+#include <cstddef>
+#include <cstdint>
+#include <tuple>
+
+using namespace arm_compute::misc::shape_calculator;
+
+namespace arm_compute
+{
+namespace
+{
+using ElementsProcessed = Steps;
+
+Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                          const GEMMReshapeInfo &gemm_info)
+{
+    ARM_COMPUTE_UNUSED(alpha);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");
+    ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16);
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");
+
+    const int m = gemm_info.m();
+    const int n = gemm_info.n();
+    const int k = gemm_info.k();
+
+    TensorShape tensor_shape0{ input0->tensor_shape() };
+    tensor_shape0.set(0, k);
+    tensor_shape0.set(1, m);
+
+    TensorShape tensor_shape1{ input1->tensor_shape() };
+    tensor_shape1.set(0, n);
+    tensor_shape1.set(1, k);
+
+    const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
+    const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
+
+    const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info0);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
+
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                                                        const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed)
+{
+    unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
+    unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
+    bool          reinterpret_output_as_3d            = (gemm_info.depth_output_gemm3d() != 0);
+
+    Window win{};
+    Window win_out{};
+    bool   window_changed = false;
+
+    // In case both input and output have to be reinterpreted as 3D tensors,
+    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)));
+
+    TensorInfo tmp_info(*output);
+
+    if(reinterpret_output_as_3d)
+    {
+        // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
+        // the window needs to be constructed on the 2D collapsed version of the tensor
+        TensorShape tmp_shape(output->tensor_shape());
+        tmp_shape.collapse(2U, 1U);
+        tmp_info.set_tensor_shape(tmp_shape);
+    }
+
+    // Configure kernel window
+    num_elems_processed_per_iteration_x = rhs_info.n0;
+    num_elems_processed_per_iteration_y = lhs_info.m0;
+
+    // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor
+    // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic
+    const int m          = gemm_info.m();
+    const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y;
+
+    win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+    win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+    AccessWindowStatic input0_access(input0, 0, 0,
+                                     input0->dimension(0),
+                                     input0->dimension(1) + bottom_pad);
+    AccessWindowStatic input1_access(input1, 0, 0,
+                                     input1->dimension(0),
+                                     input1->dimension(1));
+    AccessWindowStatic output_access(output, 0, 0,
+                                     ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x),
+                                     output->dimension(1) + bottom_pad);
+
+    window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
+                     update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor
+
+    output_access.set_valid_region(win_out, 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;
+    const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
+    collapsed                                = win.collapse(win, dimension_to_collapse);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, collapsed);
+}
+} // namespace
+
+CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::CLGEMMMatrixMultiplyReshapedOnlyRHSKernel()
+    : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false)
+{
+}
+
+void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, const GEMMLHSMatrixInfo &lhs_info,
+                                                          const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), alpha, lhs_info, rhs_info, gemm_info));
+
+    _input0                   = input0;
+    _input1                   = input1;
+    _output                   = output;
+    _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d();
+    _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
+
+    // In case both input and output have to be reinterpreted as 3D tensors,
+    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+
+    // Check if we need to slide the matrix B
+    const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
+    _slide_matrix_b                          = (_input1->info()->num_dimensions() >= num_dimensions_input0);
+
+    ElementsProcessed num_elements_processed{};
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
+    // Create build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
+    build_opts.add_option_if(std::abs(1.0f - alpha) > 0.00001f, "-DALPHA=" + float_to_string_with_full_precision(alpha));
+    build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
+    build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
+    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
+    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));
+    build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
+    build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
+    build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k()));
+    build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
+    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));
+
+
+    std::string kernel_name("gemm_mm_reshaped_only_rhs_");
+    kernel_name += rhs_info.transpose ? "t" : "nt";
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+
+    // Set config_id for enabling LWS tuning
+    _config_id = kernel_name;
+    _config_id += "_";
+    _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
+    _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
+    _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(gemm_info.k());
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(2));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(lhs_info.m0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(rhs_info.n0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(rhs_info.k0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(rhs_info.h0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(rhs_info.interleave);
+}
+
+Status CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info,
+                                                           const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
+{
+    ElementsProcessed num_elements_processed{};
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, alpha, lhs_info, rhs_info, gemm_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
+                                                              input1->clone().get(),
+                                                              output->clone().get(),
+                                                              lhs_info,
+                                                              rhs_info,
+                                                              gemm_info,
+                                                              num_elements_processed)
+                                .first);
+
+    return Status{};
+}
+
+void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    if(_input1->info()->num_dimensions() < 3)
+    {
+        // The stride_z for matrix B must be zero if we do not slice
+        ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
+    }
+
+    Window slice          = window.first_slice_window_3D();
+    Window slice_matrix_b = slice;
+
+    slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
+    slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+    if(_reinterpret_input_as_3d)
+    {
+        // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
+        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3;
+        const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom;
+        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
+    }
+
+    if(_reinterpret_output_as_3d)
+    {
+        // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
+        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
+        const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
+        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
+    }
+
+    do
+    {
+        Window slice_b = slice;
+        // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
+        // This scenario can happen when the matrix multiplication is used to perform a convolution operation
+        if(!_slide_matrix_b)
+        {
+            slice_b = slice_matrix_b;
+        }
+
+        unsigned int idx = 0;
+        add_2D_tensor_argument(idx, _input0, slice);
+        add_2D_tensor_argument(idx, _input1, slice_b);
+        add_2D_tensor_argument(idx, _output, slice);
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(window.slide_window_slice_3D(slice));
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
new file mode 100644
index 0000000000..cbbc5922dd
--- /dev/null
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
@@ -0,0 +1,349 @@
+/*
+ * Copyright (c) 2019 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/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/CL/Helper.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/GEMMFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+using namespace arm_compute::misc::shape_calculator;
+
+// Create function for CLGEMMReshapeRHSMatrixKernel
+using CLGEMMReshapeRHSMatrix = CLSynthetizeFunction<CLGEMMReshapeRHSMatrixKernel>;
+
+// Create function for CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
+using CLGEMMMatrixMultiplyReshapedOnlyRHS = CLSynthetizeFunction<CLGEMMMatrixMultiplyReshapedOnlyRHSKernel>;
+
+// Fixture for CLGEMMMatrixMultiplyReshapedOnlyRHS
+template <typename T>
+using CLGEMMMatrixMultiplyReshapedOnlyRHSFixture = GEMMMatrixMultiplyReshapedOnlyRHSValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshapedOnlyRHS>;
+
+// Fixture for CLGEMMMatrixMultiplyReshapedOnlyRHS3D
+template <typename T>
+using CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture = GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshapedOnlyRHS>;
+
+namespace
+{
+// *INDENT-OFF*
+// clang-format off
+RelativeTolerance<float> rel_tolerance_f32(0.001f);
+constexpr float          abs_tolerance_f32(0.0001f);
+
+RelativeTolerance<half> rel_tolerance_f16(half(0.2));
+constexpr float         tolerance_num_f16 = 0.02f;
+
+/** Alpha values to test - Precommit */
+const auto a_values = framework::dataset::make("alpha", {1.0f, -0.75f} );
+
+/** M values to test */
+const auto m_values = framework::dataset::make("M", 37);
+
+/** M_W values to test */
+const auto m_w_values = framework::dataset::make("M_W", 5);
+
+/** M_H values to test */
+const auto m_h_values = framework::dataset::make("M_H", 7);
+
+/** N values to test */
+const auto n_values = framework::dataset::make("N", 51);
+
+/** K values to test */
+const auto k_values = framework::dataset::make("K", 23);
+
+/** Batch size values to test */
+const auto b_values = framework::dataset::make("batch_size", 1, 3);
+
+/** M0 values to test - Precommit */
+const auto m0_values_precommit = framework::dataset::make("M0", {4, 6});
+
+/** N0 values to test - Precommit */
+const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 });
+
+/** K0 values to test - Precommit */
+const auto k0_values_precommit = framework::dataset::make("K0", { 4 });
+
+/** H0 values to test - Precommit */
+const auto h0_values_precommit = framework::dataset::make("H0", 1, 3);
+
+/** M0 values to test - Nightly */
+const auto m0_values_nightly = framework::dataset::make("M0", 2, 8);
+
+/** N0 values to test - Nightly */
+const auto n0_values_nightly = framework::dataset::make("N0", { 2, 3, 4, 8 });
+
+/** K0 values to test - Nightly */
+const auto k0_values_nightly = framework::dataset::make("K0", { 2, 3, 4, 8 });
+
+/** H0 values to test - Nightly */
+const auto h0_values_nightly = framework::dataset::make("H0", 1, 4);
+
+/** Interleave values to test with RHS matrix */
+const auto i_values_rhs = framework::dataset::make("interleave_rhs", { true, false });
+
+/** Transpose values to test with RHS matrix */
+const auto t_values_rhs = framework::dataset::make("transpose_rhs", { true });
+
+/** Configuration test */
+void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, unsigned int h0_value, bool i_value_rhs, DataType data_type)
+{
+    const unsigned int M = m_value;
+    const unsigned int N = n_value;
+    const unsigned int K = k_value;
+
+    GEMMLHSMatrixInfo lhs_info;
+    lhs_info.m0         = m0_value;
+    lhs_info.k0         = k0_value;
+
+    GEMMRHSMatrixInfo rhs_info;
+    rhs_info.n0         = n0_value;
+    rhs_info.k0         = k0_value;
+    rhs_info.h0         = h0_value;
+    rhs_info.interleave = i_value_rhs;
+    rhs_info.transpose  = true;
+
+    GEMMReshapeInfo gemm_info(M, N, K);
+
+    const TensorShape lhs_shape(K, M, b_value);
+    const TensorShape rhs_shape(N, K, b_value);
+    const TensorShape rhs_shape_reshaped = compute_rhs_reshaped_shape(TensorInfo(rhs_shape, 1, data_type),
+                                                                      rhs_info);
+
+    const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape, 1, data_type),
+                                                   TensorInfo(rhs_shape_reshaped, 1, data_type),
+                                                   gemm_info);
+
+    // Create tensors
+    CLTensor lhs          = create_tensor<CLTensor>(lhs_shape, data_type);
+    CLTensor rhs_reshaped = create_tensor<CLTensor>(rhs_shape_reshaped, data_type);
+    CLTensor dst          = create_tensor<CLTensor>(dst_shape, data_type);
+
+    ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    CLGEMMMatrixMultiplyReshapedOnlyRHS gemm;
+    gemm.configure(&lhs, &rhs_reshaped, &dst, 1.0f, lhs_info, rhs_info, gemm_info);
+}
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(GEMMMatrixMultiplyReshapedOnlyRHS)
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   framework::dataset::make("batch_size", 1)),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_rhs),
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, h0_value, i_value_rhs)
+{
+    validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, h0_value, i_value_rhs, DataType::F32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<float>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_rhs),
+                                                                   t_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   a_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<float>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_nightly),
+                                                                   n0_values_nightly),
+                                                                   k0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_rhs),
+                                                                   t_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   a_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<float>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_rhs),
+                                                                   t_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   a_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<float>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_nightly),
+                                                                   n0_values_nightly),
+                                                                   k0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_rhs),
+                                                                   t_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   a_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+TEST_SUITE_END() // FP32
+
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<half>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_rhs),
+                                                                   t_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   a_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedOnlyRHSFixture<half>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_nightly),
+                                                                   n0_values_nightly),
+                                                                   k0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_rhs),
+                                                                   t_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   a_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<half>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_rhs),
+                                                                   t_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   a_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture<half>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_nightly),
+                                                                   n0_values_nightly),
+                                                                   k0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_rhs),
+                                                                   t_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   a_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+TEST_SUITE_END() // FP16
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // GEMMMatrixMulipltyReshapedOnlyRHS
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index a6a3b67785..77d2ca61fb 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -383,6 +383,220 @@ protected:
     TensorType      _target{};
     SimpleTensor<T> _reference{};
 };
+
+template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMMatrixMultiplyReshapedOnlyRHSValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0,
+               bool interleave_rhs, bool transpose_rhs, DataType data_type, float alpha)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0 = m0;
+        lhs_info.k0 = k0;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0         = n0;
+        rhs_info.k0         = k0;
+        rhs_info.h0         = h0;
+        rhs_info.interleave = interleave_rhs;
+        rhs_info.transpose  = transpose_rhs;
+
+        // Set the tensor shapes for LHS and RHS matrices
+        const TensorShape lhs_shape(k, m, batch_size);
+        const TensorShape rhs_shape(n, k, batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type, alpha);
+        _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+
+        // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0)
+        std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity());
+        library->fill_borders_with_garbage(tensor, distribution_inf, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, DataType data_type, float alpha)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType rhs_reshaped;
+        TensorType dst;
+
+        const unsigned int M = lhs_shape[1];
+        const unsigned int N = rhs_shape[0];
+        const unsigned int K = lhs_shape[0];
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        ReshapeRHSFunctionType reshape_rhs;
+        GEMMFunctionType       gemm;
+        reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
+        gemm.configure(&lhs, &rhs_reshaped, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K));
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        rhs_reshaped.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(lhs), 0);
+        fill(AccessorType(rhs), 1);
+
+        // Compute GEMM
+        reshape_rhs.run();
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape[0]          = rhs_shape[0];
+        dst_shape[1]          = lhs_shape[1];
+
+        // Create reference
+        SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> c{ dst_shape, data_type, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+
+        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0,
+               bool interleave_rhs, bool transpose_rhs, DataType data_type, float alpha)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0 = m0;
+        lhs_info.k0 = k0;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0         = n0;
+        rhs_info.k0         = k0;
+        rhs_info.h0         = h0;
+        rhs_info.interleave = interleave_rhs;
+        rhs_info.transpose  = transpose_rhs;
+
+        // In case of GEMM3D, m is the product between m_w and m_h
+        const unsigned int m = m_w * m_h;
+
+        // Set the tensor shapes for LHS and RHS matrices
+        const TensorShape lhs_shape(k, m, batch_size);
+        const TensorShape rhs_shape(n, k, batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type, alpha, m_h);
+        _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, m_h);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, DataType data_type, float alpha,
+                              unsigned int m_h)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType rhs_reshaped;
+        TensorType dst;
+
+        const unsigned int M = lhs_shape[1];
+        const unsigned int N = rhs_shape[0];
+        const unsigned int K = lhs_shape[0];
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        ReshapeRHSFunctionType reshape_rhs;
+        GEMMFunctionType       gemm;
+        reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
+        gemm.configure(&lhs, &rhs_reshaped, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h));
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        rhs_reshaped.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(lhs), 0);
+        fill(AccessorType(rhs), 1);
+
+        // Compute GEMM
+        reshape_rhs.run();
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, unsigned int m_h)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape.set(0, rhs_shape[0]);
+        dst_shape.set(1, lhs_shape[1] / m_h);
+        dst_shape.set(2, m_h);
+        dst_shape.set(3, lhs_shape[2]);
+
+        // Create reference
+        SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> c{ dst_shape, data_type, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+
+        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute