COMPMID-1687: Optimize CLGEMMMatrixMultiplyKernel for Mali-G76 - Part1

The current implementation is limited just to FP32

Change-Id: I185ab57e483e879d7c301e9cc3033efc8b41e244
Reviewed-on: https://review.mlplatform.org/389
Reviewed-by: Anthony Barbier <Anthony.barbier@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

14 files changed, 1629 insertions, 86 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 37b92f2d6c..d89426dd32 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -78,6 +78,7 @@
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h"
 #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/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/CLGEMMMatrixMultiplyReshapedKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
new file mode 100644
index 0000000000..d0f67e6f2c
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
@@ -0,0 +1,88 @@
+/*
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H__
+#define __ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** OpenCL kernel to multiply matrices when both the input matrices LHS (input0) and RHS (input1) have been reshaped
+ *
+ * @note The input matrices @p input0 and @p input1 must be reshaped through @ref CLGEMMReshapeLHSMatrixKernel and  @ref CLGEMMReshapeRHSMatrixKernel
+ */
+class CLGEMMMatrixMultiplyReshapedKernel : public ICLKernel
+{
+public:
+    /** Default Constructor */
+    CLGEMMMatrixMultiplyReshapedKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMMatrixMultiplyReshapedKernel(const CLGEMMMatrixMultiplyReshapedKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMMatrixMultiplyReshapedKernel &operator=(const CLGEMMMatrixMultiplyReshapedKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLGEMMMatrixMultiplyReshapedKernel(CLGEMMMatrixMultiplyReshapedKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLGEMMMatrixMultiplyReshapedKernel &operator=(CLGEMMMatrixMultiplyReshapedKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  input0    Input tensor containing the LHS reshaped 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 for reshaping the input0 tensor
+     * @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 CLGEMMMatrixMultiplyReshapedKernel
+     *
+     * @param[in] input0    Input tensor containing the LHS reshaped matrix. Data type supported: F32/F16
+     * @param[in] input1    Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0
+     * @param[in] 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 for reshaping the input0 tensor
+     * @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_output_as_3d;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H__*/
\ No newline at end of file
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 33893ad877..f41d00f54d 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -619,6 +619,31 @@ inline TensorShape compute_mm_shape(const ITensorInfo &input0, const ITensorInfo
     return output_shape;
 }
 
+inline TensorShape compute_mm_shape(const ITensorInfo &input0, const ITensorInfo &input1, const GEMMReshapeInfo &gemm_info)
+{
+    ARM_COMPUTE_ERROR_ON_MSG(input0.num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4");
+
+    const bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d() != 0;
+    const int  depth_output_gemm3d      = reinterpret_output_as_3d ? gemm_info.depth_output_gemm3d() : 1;
+
+    // If the output of GEMM has to be reinterpreted as 3D, the number of input0 rows (M) is obtained collapsing the second and third
+    // dimension of the output tensor
+    const int dim0 = gemm_info.n();
+    const int dim1 = gemm_info.m() / depth_output_gemm3d;
+    const int dim2 = input0.tensor_shape()[2];
+    const int dim3 = input0.tensor_shape()[3];
+
+    TensorShape output_shape{ input0.tensor_shape() };
+
+    output_shape.set(0, dim0);
+    output_shape.set(1, dim1);
+    output_shape.set(2, reinterpret_output_as_3d ? depth_output_gemm3d : dim2);
+    output_shape.set(3, reinterpret_output_as_3d ? dim2 : dim3);
+    output_shape.set(4, reinterpret_output_as_3d ? dim3 : 1);
+
+    return output_shape;
+}
+
 inline TensorShape compute_output_stage_shape(const ITensorInfo &input, unsigned int gemm_3d_depth = 1, bool batch_size_on_z = false)
 {
     ARM_COMPUTE_ERROR_ON(input.data_layout() != DataLayout::NHWC && gemm_3d_depth > 1);
diff --git a/arm_compute/runtime/CL/functions/CLGEMM.h b/arm_compute/runtime/CL/functions/CLGEMM.h
index c4513f29d9..7d47194e56 100644
--- a/arm_compute/runtime/CL/functions/CLGEMM.h
+++ b/arm_compute/runtime/CL/functions/CLGEMM.h
@@ -27,6 +27,9 @@
 #include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
 #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/CLGEMMReshapeLHSMatrixKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h"
 #include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
@@ -39,9 +42,12 @@ class ICLTensor;
 
 /** Basic function to execute GEMM on OpenCL. This function calls the following OpenCL kernels:
  *
- *  -# @ref CLGEMMInterleave4x4Kernel (only if the reshaped GEMM is selected by the heuristic model)
- *  -# @ref CLGEMMTranspose1xWKernel (only if the reshaped GEMM is selected by the heuristic model)
- *  -# @ref CLGEMMMatrixMultiplyKernel
+ *  -# @ref CLGEMMInterleave4x4Kernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target is NOT Mali-G76)
+ *  -# @ref CLGEMMReshapeLHSMatrixKernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target IS Mali-G76)
+ *  -# @ref CLGEMMTranspose1xWKernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target is NOT Mali-G76)
+ *  -# @ref CLGEMMReshapeRHSMatrixKernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target IS Mali-G76)
+ *  -# @ref CLGEMMMatrixMultiplyKernel (if GPU target is NOT G76 or if the reshaped GEMM is NOT selected)
+ *  -# @ref CLGEMMMatrixMultiplyReshapedKernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target IS Mali-G76)
  *  -# @ref CLGEMMMatrixAdditionKernel (if c != nullptr and beta != 0.0)
  *
  */
@@ -100,18 +106,22 @@ public:
     void prepare() override;
 
 private:
-    CLMemoryGroup              _memory_group;
-    CLGEMMInterleave4x4Kernel  _interleave_kernel;
-    CLGEMMTranspose1xWKernel   _transpose_kernel;
-    CLGEMMMatrixMultiplyKernel _mm_kernel;
-    CLGEMMMatrixAdditionKernel _ma_kernel;
-    CLTensor                   _tmp_a;
-    CLTensor                   _tmp_b;
-    const ICLTensor           *_original_b;
-    bool                       _is_interleaved_transposed;
-    bool                       _run_addition;
-    bool                       _reshape_b_only_on_first_run;
-    bool                       _is_prepared;
+    CLMemoryGroup                      _memory_group;
+    CLGEMMInterleave4x4Kernel          _interleave_kernel; // TODO - COMPMID-1835: Remove this kernel and use CLGEMMReshapeLHSMatrixKernel
+    CLGEMMTranspose1xWKernel           _transpose_kernel;  // TODO - COMPMID-1836: Remove this kernel and use CLGEMMReshapeRHSMatrixKernel
+    CLGEMMMatrixMultiplyKernel         _mm_kernel;
+    CLGEMMMatrixAdditionKernel         _ma_kernel;
+    CLGEMMReshapeLHSMatrixKernel       _reshape_lhs_kernel;
+    CLGEMMReshapeRHSMatrixKernel       _reshape_rhs_kernel;
+    CLGEMMMatrixMultiplyReshapedKernel _mm_reshaped_kernel;
+    CLTensor                           _tmp_a;
+    CLTensor                           _tmp_b;
+    const ICLTensor                   *_original_b;
+    bool                               _is_interleaved_transposed;
+    bool                               _run_addition;
+    bool                               _reshape_b_only_on_first_run;
+    bool                               _is_prepared;
+    bool                               _is_G76_path; // TODO: To be removed once completed COMPMID-1835 and COMPMID-1836
 };
 }
 
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 54fc618bdf..03bc8d15db 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -280,6 +280,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemm_mm_floating_point_f16_bifrost_acc32", "gemm.cl" },
     { "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_lc_vm_f32", "gemm.cl" },
     { "gemm_transpose1xW", "gemm.cl" },
     { "gemm_reshape_lhs_matrix_nt", "gemm.cl" },
@@ -319,6 +320,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "im2col_generic_nchw", "im2col.cl" },
     { "im2col_generic_padx0_pady0_nchw", "im2col.cl" },
     { "im2col3x3_nhwc", "im2col.cl" },
+    { "im2col9x9_nhwc", "im2col.cl" },
     { "im2col_generic_nhwc", "im2col.cl" },
     { "init_level", "optical_flow_pyramid_lk.cl" },
     { "init_level_max", "optical_flow_pyramid_lk.cl" },
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index 40ee1d45ad..d37dd2d2d6 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -68,17 +68,17 @@ __kernel void gemm_reshape_lhs_matrix_nt(TENSOR3D_DECLARATION(src),
 #endif // REINTERPRET_INPUT_AS_3D
                                         )
 {
-// Block size
+    // Block size
 #define BLOCK_SIZE ((M0) * (K0))
 
-// Output offset X
+    // 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
+    // Output step X
 #if defined(INTERLEAVE)
 #define OUTPUT_STEP_X (K0) * (V0)
 #else // Do not interleave
@@ -711,27 +711,27 @@ __kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src),
     // 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);
+    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);
+    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);
+    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);
+    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
@@ -832,6 +832,471 @@ __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(V0) && defined(H0) && defined(K) && defined(DATA_TYPE)
+
+#define ARM_DOT(x, y, val)          \
+    ({                              \
+        val = fma(x.s0, y.s0, val); \
+        val = fma(x.s1, y.s1, val); \
+        val = fma(x.s2, y.s2, val); \
+        val = fma(x.s3, y.s3, val); \
+    })
+
+#if K0 == 4
+#define ARM_DOT_K0(a, b, c) \
+    ({                      \
+        ARM_DOT(a, b, c);   \
+    })
+#elif K0 == 8 // K0 == 8
+#define ARM_DOT_K0(a, b, c)               \
+    ({                                    \
+        ARM_DOT((a).s0123, (b).s0123, c); \
+        ARM_DOT((a).s4567, (b).s4567, c); \
+    })
+#elif K0 == 16 // K0 == 16
+#define ARM_DOT_K0(a, b, c)               \
+    ({                                    \
+        ARM_DOT((a).s0123, (b).s0123, c); \
+        ARM_DOT((a).s4567, (b).s4567, c); \
+        ARM_DOT((a).s89AB, (b).s89AB, c); \
+        ARM_DOT((a).sCDEF, (b).sCDEF, c); \
+    })
+#else // K0 not supported
+#error "K0 value not supported"
+#endif // K0 conditions
+
+#if N0 == 2
+#define ARM_DOT_K0XN0(a, b, c)           \
+    ({                                   \
+        ARM_DOT_K0((a), (b##0), (c.s0)); \
+        ARM_DOT_K0((a), (b##1), (c.s1)); \
+    })
+#elif N0 == 4 // N0 == 4
+#define ARM_DOT_K0XN0(a, b, c)           \
+    ({                                   \
+        ARM_DOT_K0((a), (b##0), (c.s0)); \
+        ARM_DOT_K0((a), (b##1), (c.s1)); \
+        ARM_DOT_K0((a), (b##2), (c.s2)); \
+        ARM_DOT_K0((a), (b##3), (c.s3)); \
+    })
+#elif N0 == 8 // N0 == 8
+#define ARM_DOT_K0XN0(a, b, c)           \
+    ({                                   \
+        ARM_DOT_K0((a), (b##0), (c.s0)); \
+        ARM_DOT_K0((a), (b##1), (c.s1)); \
+        ARM_DOT_K0((a), (b##2), (c.s2)); \
+        ARM_DOT_K0((a), (b##3), (c.s3)); \
+        ARM_DOT_K0((a), (b##4), (c.s4)); \
+        ARM_DOT_K0((a), (b##5), (c.s5)); \
+        ARM_DOT_K0((a), (b##6), (c.s6)); \
+        ARM_DOT_K0((a), (b##7), (c.s7)); \
+    })
+#elif N0 == 16 // N0 == 16
+#define ARM_DOT_K0XN0(a, b, c)           \
+    ({                                   \
+        ARM_DOT_K0((a), (b##0), (c.s0)); \
+        ARM_DOT_K0((a), (b##1), (c.s1)); \
+        ARM_DOT_K0((a), (b##2), (c.s2)); \
+        ARM_DOT_K0((a), (b##3), (c.s3)); \
+        ARM_DOT_K0((a), (b##4), (c.s4)); \
+        ARM_DOT_K0((a), (b##5), (c.s5)); \
+        ARM_DOT_K0((a), (b##6), (c.s6)); \
+        ARM_DOT_K0((a), (b##7), (c.s7)); \
+        ARM_DOT_K0((a), (b##8), (c.s8)); \
+        ARM_DOT_K0((a), (b##9), (c.s9)); \
+        ARM_DOT_K0((a), (b##A), (c.sA)); \
+        ARM_DOT_K0((a), (b##B), (c.sB)); \
+        ARM_DOT_K0((a), (b##C), (c.sC)); \
+        ARM_DOT_K0((a), (b##D), (c.sD)); \
+        ARM_DOT_K0((a), (b##E), (c.sE)); \
+        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 must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed
+ *  The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed
+ *
+ * @note The number of columns in the RHS matrix NOT reshaped needs to be passed at compile time using -DK (i.e. -Dk=128).
+ * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (i.e. -DM0=4, -DN0=8, -DK0=4).
+ * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (i.e. -DV0=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 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
+ *  - N0 = 2, 4, 8, 16
+ *  - K0 = 4, 8, 16
+ *
+ * @note In case the output has to be reinterpreted as a 3D tensor (i.e. output of convolution layer), the following information must be passed at compile time:
+ *       -# 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 NOT reshaped
+ *
+ * @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 src0_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: S32
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @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]  dst_cross_plane_pad               (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
+ */
+__kernel void gemm_mm_reshaped_lhs_nt_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_OUTPUT_AS_3D)
+                                            ,
+                                            uint dst_cross_plane_pad
+#endif // REINTERPRET_OUTPUT_AS_3D
+                                           )
+{
+    // Block size
+#define LHS_BLOCK_SIZE ((K0) * (M0))
+
+#if defined(LHS_INTERLEAVE)
+#define LHS_OFFSET_X (K0)
+#define LHS_STEP_X ((K0) * (V0))
+#define LHS_STEP_LOOP (1)
+#else // defined(INTERLEAVE)
+#define LHS_OFFSET_X (LHS_BLOCK_SIZE)
+#define LHS_STEP_X (K0)
+#define LHS_STEP_LOOP (V0)
+#endif // defined(INTERLEAVE)
+
+    // 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)
+
+    // Compute LHS matrix address
+    __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (get_global_id(1) % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(1) / V0) * (uint)lhs_stride_y +
+                               (get_global_id(2) * lhs_stride_z);
+
+    // Compute RHS matrix address
+    __global uchar *rhs_addr = rhs_ptr + rhs_offset_first_element_in_bytes + (get_global_id(0) % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(0) / (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_addr += (get_global_id(2) % MATRIX_B_DEPTH) * rhs_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
+    rhs_addr += get_global_id(2) * rhs_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    // Initialize the accumulators
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    c0 = 0;
+#if M0 > 1
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    c1 = 0;
+#endif // M0 > 1
+#if M0 > 2
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    c2 = 0;
+#endif // M0 > 2
+#if M0 > 3
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    c3 = 0;
+#endif // M0 > 3
+#if M0 > 4
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    c4 = 0;
+#endif // M0 > 4
+#if M0 > 5
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    c5 = 0;
+#endif // M0 > 5
+#if M0 > 6
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    c6 = 0;
+#endif // M0 > 6
+#if M0 > 7
+    VEC_DATA_TYPE(DATA_TYPE, N0)
+    c7 = 0;
+#endif // M0 > 7
+
+    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
+        // 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)));
+#if M0 > 1
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a1 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_addr + 1 * LHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 1
+#if M0 > 2
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a2 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_addr + 2 * LHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 2
+#if M0 > 3
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a3 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_addr + 3 * LHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 3
+#if M0 > 4
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a4 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_addr + 4 * LHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 4
+#if M0 > 5
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a5 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_addr + 5 * LHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 5
+#if M0 > 6
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a6 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_addr + 6 * LHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 6
+#if M0 > 7
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        a7 = VLOAD(K0)(0, (__global DATA_TYPE *)(lhs_addr + 7 * LHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 7
+
+        // Load values from RHS matrix
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b0 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b1 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 1 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#if N0 > 2
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b2 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 2 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b3 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 3 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 2
+#if N0 > 4
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b4 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 4 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b5 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 5 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b6 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 6 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b7 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 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_addr + 8 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        b9 = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 9 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bA = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 10 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bB = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 11 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bC = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 12 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bD = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 13 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bE = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 14 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VEC_DATA_TYPE(DATA_TYPE, K0)
+        bF = VLOAD(K0)(0, (__global DATA_TYPE *)(rhs_addr + 15 * RHS_STEP_X * sizeof(DATA_TYPE)));
+#endif // N0 > 8
+
+        // Accumulate
+        ARM_DOT_K0XN0(a0, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(a1, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(a2, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(a3, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(a4, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(a5, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(a6, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(a7, b, c7);
+#endif // M0 > 7
+
+        lhs_addr += (M0 * LHS_STEP_X * LHS_STEP_LOOP) * sizeof(DATA_TYPE);
+        rhs_addr += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE);
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * dst_stride_y);
+
+    uint zout0 = 0;
+    uint zout1 = 0;
+    uint zout2 = 0;
+    uint zout3 = 0;
+    uint zout4 = 0;
+    uint zout5 = 0;
+    uint zout6 = 0;
+    uint 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 (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    zout0 = (0 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout0 = min((uint)(DEPTH_GEMM3D - 1), zout0);
+    zout0 *= (dst_cross_plane_pad * dst_stride_z);
+#if M0 > 1
+    zout1 = (1 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout1 = min((uint)(DEPTH_GEMM3D - 1), zout1);
+    zout1 *= (dst_cross_plane_pad * dst_stride_z);
+#endif // M0 > 1
+#if M0 > 2
+    zout2 = (2 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout2 = min((uint)(DEPTH_GEMM3D - 1), zout2);
+    zout2 *= (dst_cross_plane_pad * dst_stride_z);
+#endif // M0 > 2
+#if M0 > 3
+    zout3 = (3 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout3 = min((uint)(DEPTH_GEMM3D - 1), zout3);
+    zout3 *= (dst_cross_plane_pad * dst_stride_z);
+#endif // M0 > 3
+#if M0 > 4
+    zout4 = (4 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout4 = min((uint)(DEPTH_GEMM3D - 1), zout4);
+    zout4 *= (dst_cross_plane_pad * dst_stride_z);
+#endif // M0 > 4
+#if M0 > 5
+    zout5 = (5 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout5 = min((uint)(DEPTH_GEMM3D - 1), zout5);
+    zout5 *= (dst_cross_plane_pad * dst_stride_z);
+#endif // M0 > 5
+#if M0 > 6
+    zout6 = (6 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout6 = min((uint)(DEPTH_GEMM3D - 1), zout6);
+    zout6 *= (dst_cross_plane_pad * dst_stride_z);
+#endif // M0 > 6
+#if M0 > 6
+    zout7 = (7 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D;
+    zout7 = min((uint)(DEPTH_GEMM3D - 1), zout7);
+    zout7 *= (dst_cross_plane_pad * dst_stride_z);
+#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 += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D;
+
+#else // defined(REINTERPRET_OUTPUT_AS_3D)
+
+    // Add offset for batched GEMM
+    dst_addr += get_global_id(2) * 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 LHS_BLOCK_SIZE
+#undef LHS_OFFSET_X
+#undef LHS_STEP_X
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+}
+#endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(K) && defined(DATA_TYPE)
+
 #if defined(TRANSPOSE_W) && defined(MULT_TRANSPOSE1XW_WIDTH)
 
 #if ELEMENT_SIZE == 1
diff --git a/src/core/CL/cl_kernels/im2col.cl b/src/core/CL/cl_kernels/im2col.cl
index 186d5a80ad..2bf59e4a99 100644
--- a/src/core/CL/cl_kernels/im2col.cl
+++ b/src/core/CL/cl_kernels/im2col.cl
@@ -1029,6 +1029,177 @@ __kernel void im2col3x3_nhwc(
 #endif // HAS_BIAS
 }
 
+#if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
+#define IM2COL1x9(i)                                                                                                                                                       \
+    ({                                                                                                                                                                     \
+        yi_coord = yi - (int)PAD_TOP + i * DILATION_Y;                                                                                                                     \
+        yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));                                                                                                            \
+        \
+        offset0 = xi_offset0 + (yi_coord * (int)src_stride_z);                                                                                                             \
+        offset1 = xi_offset1 + (yi_coord * (int)src_stride_z);                                                                                                             \
+        \
+        VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s0));                                                                          \
+        VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s1));                                                                          \
+        VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s2));                                                                          \
+        VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s3));                                                                          \
+        VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s4));                                                                          \
+        VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s5));                                                                          \
+        VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s6));                                                                          \
+        VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s7));                                                                          \
+        VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset1));                                                                             \
+        \
+        int y_cond = (int)((uint)(yi - (int)PAD_TOP + i * DILATION_Y) >= (uint)(SRC_HEIGHT));                                                                              \
+        values0    = select(values0, (VECTOR_N)PAD_VALUE, (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))y_cond || (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))(x_cond0.s0)); \
+        values1    = select(values1, (VECTOR_N)PAD_VALUE, (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))y_cond || (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))(x_cond0.s1)); \
+        values2    = select(values2, (VECTOR_N)PAD_VALUE, (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))y_cond || (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))(x_cond0.s2)); \
+        values3    = select(values3, (VECTOR_N)PAD_VALUE, (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))y_cond || (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))(x_cond0.s3)); \
+        values4    = select(values4, (VECTOR_N)PAD_VALUE, (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))y_cond || (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))(x_cond0.s4)); \
+        values5    = select(values5, (VECTOR_N)PAD_VALUE, (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))y_cond || (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))(x_cond0.s5)); \
+        values6    = select(values6, (VECTOR_N)PAD_VALUE, (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))y_cond || (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))(x_cond0.s6)); \
+        values7    = select(values7, (VECTOR_N)PAD_VALUE, (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))y_cond || (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))(x_cond0.s7)); \
+        values8    = select(values8, (VECTOR_N)PAD_VALUE, (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))y_cond || (VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE))(x_cond1));    \
+        \
+        VSTORE(VECTOR_SIZE)                                                                                                                                                \
+        (values0, 0, (__global DATA_TYPE *)(output_ptr) + (0 + i * 9) * SRC_DEPTH);                                                                                        \
+        VSTORE(VECTOR_SIZE)                                                                                                                                                \
+        (values1, 0, (__global DATA_TYPE *)(output_ptr) + (1 + i * 9) * SRC_DEPTH);                                                                                        \
+        VSTORE(VECTOR_SIZE)                                                                                                                                                \
+        (values2, 0, (__global DATA_TYPE *)(output_ptr) + (2 + i * 9) * SRC_DEPTH);                                                                                        \
+        VSTORE(VECTOR_SIZE)                                                                                                                                                \
+        (values3, 0, (__global DATA_TYPE *)(output_ptr) + (3 + i * 9) * SRC_DEPTH);                                                                                        \
+        VSTORE(VECTOR_SIZE)                                                                                                                                                \
+        (values4, 0, (__global DATA_TYPE *)(output_ptr) + (4 + i * 9) * SRC_DEPTH);                                                                                        \
+        VSTORE(VECTOR_SIZE)                                                                                                                                                \
+        (values5, 0, (__global DATA_TYPE *)(output_ptr) + (5 + i * 9) * SRC_DEPTH);                                                                                        \
+        VSTORE(VECTOR_SIZE)                                                                                                                                                \
+        (values6, 0, (__global DATA_TYPE *)(output_ptr) + (6 + i * 9) * SRC_DEPTH);                                                                                        \
+        VSTORE(VECTOR_SIZE)                                                                                                                                                \
+        (values7, 0, (__global DATA_TYPE *)(output_ptr) + (7 + i * 9) * SRC_DEPTH);                                                                                        \
+        VSTORE(VECTOR_SIZE)                                                                                                                                                \
+        (values8, 0, (__global DATA_TYPE *)(output_ptr) + (8 + i * 9) * SRC_DEPTH);                                                                                        \
+    })
+#else // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
+#define IM2COL1x9(i)                                                                              \
+    ({                                                                                            \
+        yi_coord = yi - (int)PAD_TOP + i * DILATION_Y;                                            \
+        yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));                                   \
+        \
+        offset0 = xi_offset0 + (yi_coord * (int)src_stride_z);                                    \
+        offset1 = xi_offset1 + (yi_coord * (int)src_stride_z);                                    \
+        \
+        VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s0)); \
+        VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s1)); \
+        VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s2)); \
+        VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s3)); \
+        VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s4)); \
+        VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s5)); \
+        VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s6)); \
+        VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s7)); \
+        VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset1));    \
+        \
+        VSTORE(VECTOR_SIZE)                                                                       \
+        (values0, 0, (__global DATA_TYPE *)(output_ptr) + (0 + i * 9) * SRC_DEPTH);               \
+        VSTORE(VECTOR_SIZE)                                                                       \
+        (values1, 0, (__global DATA_TYPE *)(output_ptr) + (1 + i * 9) * SRC_DEPTH);               \
+        VSTORE(VECTOR_SIZE)                                                                       \
+        (values2, 0, (__global DATA_TYPE *)(output_ptr) + (2 + i * 9) * SRC_DEPTH);               \
+        VSTORE(VECTOR_SIZE)                                                                       \
+        (values3, 0, (__global DATA_TYPE *)(output_ptr) + (3 + i * 9) * SRC_DEPTH);               \
+        VSTORE(VECTOR_SIZE)                                                                       \
+        (values4, 0, (__global DATA_TYPE *)(output_ptr) + (4 + i * 9) * SRC_DEPTH);               \
+        VSTORE(VECTOR_SIZE)                                                                       \
+        (values5, 0, (__global DATA_TYPE *)(output_ptr) + (5 + i * 9) * SRC_DEPTH);               \
+        VSTORE(VECTOR_SIZE)                                                                       \
+        (values6, 0, (__global DATA_TYPE *)(output_ptr) + (6 + i * 9) * SRC_DEPTH);               \
+        VSTORE(VECTOR_SIZE)                                                                       \
+        (values7, 0, (__global DATA_TYPE *)(output_ptr) + (7 + i * 9) * SRC_DEPTH);               \
+        VSTORE(VECTOR_SIZE)                                                                       \
+        (values8, 0, (__global DATA_TYPE *)(output_ptr) + (8 + i * 9) * SRC_DEPTH);               \
+    })
+#endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
+
+/** This kernel performs im2col when the kernel size is 9x9 and the data layout is NHWC
+ *
+ * @note This kernel computes VECTOR_SIZE elements
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
+ * @note The kernel depth must be passed at compile time using -DSRC_DEPTH: e.g. -DSRC_DEPTH=3
+ * @note The stride along the Y direction must be passed at compile time using -DSTRIDE_Y: e.g. -DSTRIDE_Y=1
+ * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QASYMM8/F16/F32
+ * @param[in]  src_stride_x                      Stride of the source 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 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 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 tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination tensor 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 tensor 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 tensor
+ * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes).
+ * @param[in]  dst_stride_w                      Stride of the destination tensor in W dimension (in bytes).
+ */
+__kernel void im2col9x9_nhwc(
+    TENSOR3D_DECLARATION(src),
+    IMAGE_DECLARATION(dst),
+    uint src_stride_w,
+    uint dst_stride_w)
+{
+    const int ch    = min((int)(get_global_id(0) * VECTOR_SIZE), LAST_ACCESSED); // input feature map
+    const int yo    = get_global_id(1);
+    const int batch = get_global_id(2); // batch size
+
+    // Calculate input indices
+    const int xi = (get_global_id(1) % CONVOLVED_WIDTH) * STRIDE_X;
+    const int yi = (get_global_id(1) / (int)CONVOLVED_WIDTH) * STRIDE_Y;
+
+    // Get input and output address
+    __global uchar *input_ptr  = src_ptr + src_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + batch * (int)src_stride_w;
+    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + yo * (int)dst_stride_y + batch * (int)dst_stride_w;
+
+    int  yi_coord = 0;
+    int8 offset0  = 0;
+    int  offset1  = 0;
+
+    // Clamp xi
+    int8 xi_offset0 = ((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT);
+    int  xi_offset1 = ((int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT);
+
+#if PAD_TOP != 0 || PAD_BOTTOM != 0
+#define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)
+    xi_offset0 = CLAMP(xi_offset0, (int8)0, (int8)(SRC_WIDTH - 1));
+    xi_offset1 = CLAMP(xi_offset1, (int)0, (int)(SRC_WIDTH - 1));
+#endif // PAD_TOP != 0 || PAD_BOTTOM != 0
+    xi_offset0 *= (int8)src_stride_y;
+    xi_offset1 *= (int)src_stride_y;
+
+    // Out-of-bound condition for X
+    int8 x_cond0 = (((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT) < (int8)0) || (((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT) >= (int8)SRC_WIDTH);
+    int  x_cond1 = (((int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT) < (int)0) || (((int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT) >= (int)SRC_WIDTH);
+
+    IM2COL1x9(0);
+    IM2COL1x9(1);
+    IM2COL1x9(2);
+    IM2COL1x9(3);
+    IM2COL1x9(4);
+    IM2COL1x9(5);
+    IM2COL1x9(6);
+    IM2COL1x9(7);
+    IM2COL1x9(8);
+
+#ifdef HAS_BIAS
+    if((ch + VECTOR_SIZE) >= SRC_DEPTH)
+    {
+        *((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * 81) = 1.0f;
+    }
+#endif // HAS_BIAS
+}
+
 /** This opencl kernel performs a generic im2col implementation when the data layout is NHWC
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp
new file mode 100644
index 0000000000..1ecde3e558
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp
@@ -0,0 +1,308 @@
+/*
+ * 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/CLGEMMMatrixMultiplyReshapedKernel.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;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace arm_compute
+{
+class Coordinates;
+} // 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(lhs_info.transpose);
+    ARM_COMPUTE_RETURN_ERROR_ON(!rhs_info.transpose);
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0);
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);
+
+    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_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));
+    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_info_reshaped0);
+    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;
+
+    // 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,
+                                     ceil_to_multiple(input0->dimension(0), num_elems_processed_per_iteration_y),
+                                     input0->dimension(1));
+    AccessWindowStatic input1_access(input1, 0, 0,
+                                     ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x),
+                                     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
+
+CLGEMMMatrixMultiplyReshapedKernel::CLGEMMMatrixMultiplyReshapedKernel()
+    : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false)
+{
+}
+
+void CLGEMMMatrixMultiplyReshapedKernel::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_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
+
+    // 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_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
+    build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
+    build_opts.add_option_if(_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(lhs_info.interleave, "-DLHS_INTERLEAVE");
+    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(lhs_info.k0));
+    build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));
+    build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
+
+    std::string kernel_name("gemm_mm_reshaped_");
+    kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";
+    kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_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_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(lhs_info.k0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(lhs_info.v0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(rhs_info.h0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(lhs_info.interleave);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(rhs_info.interleave);
+}
+
+Status CLGEMMMatrixMultiplyReshapedKernel::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 CLGEMMMatrixMultiplyReshapedKernel::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_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;
+        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));
+}
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLIm2ColKernel.cpp b/src/core/CL/kernels/CLIm2ColKernel.cpp
index 54ef23f2a2..e3d8df53e5 100644
--- a/src/core/CL/kernels/CLIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLIm2ColKernel.cpp
@@ -192,11 +192,15 @@ Im2ColConfiguration configure_opencl_kernel(const ITensorInfo *input, const Size
         num_elems_processed_per_iteration = 2;
         is_padding_required_nchw          = false;
 
-        // Only the 3x3 case is optimized for NHWC
+        // Only the 3x3 and 9x9 cases are optimized for NHWC
         if(kernel_dims == Size2D(3U, 3U))
         {
             kernel_name = "im2col3x3_";
         }
+        else if(kernel_dims == Size2D(9U, 9U))
+        {
+            kernel_name = "im2col9x9_";
+        }
 
         build_opts.add_option("-DVECTOR_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
         build_opts.add_option("-DLAST_ACCESSED=" + support::cpp11::to_string(std::max(static_cast<int>(input_channel - num_elems_processed_per_iteration), 0)));
diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp
index baa0cf46dc..d0db8766d9 100644
--- a/src/runtime/CL/functions/CLGEMM.cpp
+++ b/src/runtime/CL/functions/CLGEMM.cpp
@@ -40,25 +40,32 @@ using namespace arm_compute::misc::shape_calculator;
 
 namespace
 {
-inline bool is_interleaved_transposed(int m, int n, int k, DataType data_type, bool reshape_b_only_on_first_run, GPUTarget gpu_target)
+inline bool is_interleaved_transposed(unsigned int m, unsigned int n, unsigned int k, DataType data_type, bool reshape_b_only_on_first_run, GPUTarget gpu_target)
 {
     bool flag = true;
 
     if(gpu_target_is_in(gpu_target, GPUTarget::G52, GPUTarget::G52LIT, GPUTarget::G71, GPUTarget::G72, GPUTarget::G76))
     {
-        // COMPMID-852
-        if(k > 256 && m > 4 && is_data_type_float(data_type) && reshape_b_only_on_first_run)
+        if((m > 1) && n < 16)
         {
-            constexpr float alpha = 3.2f;
-            constexpr float fact0 = 1.51f;
-            constexpr float fact1 = 1.66f;
-            constexpr float ops   = 12.0f;
-            const float     scale = k > 1024 ? 1.07f : 1.0f;
-            flag                  = alpha + ((n * fact0) / ops) < ((fact1 * n * scale) / ops);
+            flag = true;
         }
         else
         {
-            flag = false;
+            // COMPMID-852
+            if(k > 256 && m > 4 && is_data_type_float(data_type) && reshape_b_only_on_first_run)
+            {
+                constexpr float alpha = 3.2f;
+                constexpr float fact0 = 1.51f;
+                constexpr float fact1 = 1.66f;
+                constexpr float ops   = 12.0f;
+                const float     scale = k > 1024 ? 1.07f : 1.0f;
+                flag                  = alpha + ((n * fact0) / ops) < ((fact1 * n * scale) / ops);
+            }
+            else
+            {
+                flag = false;
+            }
         }
     }
     else
@@ -69,6 +76,43 @@ inline bool is_interleaved_transposed(int m, int n, int k, DataType data_type, b
 
     return flag;
 }
+
+inline void select_gemm_configuration(unsigned int m, unsigned int n, GEMMLHSMatrixInfo &lhs_info, GEMMRHSMatrixInfo &rhs_info)
+{
+    // Heuristic selection for GEMM
+    if(n <= 4)
+    {
+        // Configure GEMMLHSMatrixInfo
+        lhs_info.m0         = 4;
+        lhs_info.k0         = 8;
+        lhs_info.v0         = lhs_info.m0 * 16 < m ? 2 : 16;
+        lhs_info.interleave = true;
+        lhs_info.transpose  = false;
+
+        // Configure GEMMRHSMatrixInfo
+        rhs_info.n0         = 2;
+        rhs_info.k0         = lhs_info.k0;
+        rhs_info.h0         = rhs_info.n0 * 16 < n ? 2 : 16;
+        rhs_info.interleave = false;
+        rhs_info.transpose  = true;
+    }
+    else
+    {
+        // Configure GEMMLHSMatrixInfo
+        lhs_info.m0         = (m * n) / 24 > 2048 ? 6 : 5;
+        lhs_info.k0         = 4;
+        lhs_info.v0         = 32;
+        lhs_info.interleave = false;
+        lhs_info.transpose  = false;
+
+        // Configure GEMMRHSMatrixInfo
+        rhs_info.n0         = 4;
+        rhs_info.k0         = lhs_info.k0;
+        rhs_info.h0         = 32;
+        rhs_info.interleave = true;
+        rhs_info.transpose  = true;
+    }
+}
 } // namespace
 
 CLGEMM::CLGEMM(std::shared_ptr<IMemoryManager> memory_manager)
@@ -77,13 +121,17 @@ CLGEMM::CLGEMM(std::shared_ptr<IMemoryManager> memory_manager)
       _transpose_kernel(),
       _mm_kernel(),
       _ma_kernel(),
+      _reshape_lhs_kernel(),
+      _reshape_rhs_kernel(),
+      _mm_reshaped_kernel(),
       _tmp_a(),
       _tmp_b(),
       _original_b(nullptr),
       _is_interleaved_transposed(false),
       _run_addition(false),
       _reshape_b_only_on_first_run(false),
-      _is_prepared(false)
+      _is_prepared(false),
+      _is_G76_path(false)
 {
 }
 
@@ -112,13 +160,14 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *
     // Arguments used by GEMMReshapeInfo
     // If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo
     // in order to know how the matrices have been reshaped
-    bool      reinterpret_input_as_3d   = gemm_info.reinterpret_input_as_3d();
-    const int m                         = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
-    const int n                         = b->info()->dimension(0);
-    const int k                         = a->info()->dimension(0);
-    const int depth_output_gemm3d       = gemm_info.depth_output_gemm3d();
-    int       mult_transpose1xW_width   = 1;
-    int       mult_interleave4x4_height = 1;
+    DataType           data_type                 = a->info()->data_type();
+    bool               reinterpret_input_as_3d   = gemm_info.reinterpret_input_as_3d();
+    const unsigned int m                         = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
+    const unsigned int n                         = b->info()->dimension(0);
+    const unsigned int k                         = a->info()->dimension(0);
+    const int          depth_output_gemm3d       = gemm_info.depth_output_gemm3d();
+    int                mult_transpose1xW_width   = 1;
+    int                mult_interleave4x4_height = 1;
 
     if(get_arch_from_target(gpu_target) == GPUTarget::BIFROST)
     {
@@ -129,6 +178,10 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *
     // Check if we need to reshape the matrix A and matrix B
     _is_interleaved_transposed = is_interleaved_transposed(m, n, k, a->info()->data_type(), _reshape_b_only_on_first_run, gpu_target);
 
+    // Check if we can run the new reshaped GEMM
+    _is_G76_path = (gpu_target == GPUTarget::G76) && _is_interleaved_transposed && (data_type == DataType::F32);
+    ;
+
     // if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
     if(_is_interleaved_transposed)
     {
@@ -145,19 +198,40 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *
         }
         // _tmp_a and _tmp_b will be auto configured in _interleave_kernel and in _transpose_kernel
 
-        // Configure interleave kernel
-        _interleave_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d());
+        if(_is_G76_path)
+        {
+            GEMMLHSMatrixInfo lhs_info;
+            GEMMRHSMatrixInfo rhs_info;
+
+            // Pick up the GEMM configuration based on M,N and K
+            select_gemm_configuration(m, n, lhs_info, rhs_info);
 
-        // Configure transpose kernel
-        _transpose_kernel.configure(b, &_tmp_b, mult_transpose1xW_width);
+            _reshape_lhs_kernel.configure(a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d());
+            _reshape_rhs_kernel.configure(b, &_tmp_b, rhs_info);
+
+            // Configure and tune matrix multiply kernel
+            _mm_reshaped_kernel.configure(matrix_a, matrix_b, output, alpha, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1,
+                                                                                                                 depth_output_gemm3d, reinterpret_input_as_3d));
+        }
+        else
+        {
+            // Configure interleave kernel
+            _interleave_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d());
+
+            // Configure transpose kernel
+            _transpose_kernel.configure(b, &_tmp_b, mult_transpose1xW_width);
+        }
     }
 
-    // Configure and tune matrix multiply kernel
-    _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k,
-                                                                                                        mult_transpose1xW_width, mult_interleave4x4_height,
-                                                                                                        depth_output_gemm3d, reinterpret_input_as_3d),
-                         gemm_info.fp_mixed_precision());
-    CLScheduler::get().tune_kernel_static(_mm_kernel);
+    if(!_is_G76_path)
+    {
+        // Configure and tune matrix multiply kernel
+        _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k,
+                                                                                                            mult_transpose1xW_width, mult_interleave4x4_height,
+                                                                                                            depth_output_gemm3d, reinterpret_input_as_3d),
+                             gemm_info.fp_mixed_precision());
+        CLScheduler::get().tune_kernel_static(_mm_kernel);
+    }
 
     if(_is_interleaved_transposed)
     {
@@ -197,13 +271,14 @@ Status CLGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso
     // Arguments used by GEMMReshapeInfo
     // If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo
     // in order to know how the matrices have been reshaped
-    bool      reinterpret_input_as_3d   = gemm_info.reinterpret_input_as_3d();
-    const int m                         = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
-    const int n                         = b->dimension(0);
-    const int k                         = a->dimension(0);
-    int       mult_transpose1xW_width   = 1;
-    int       mult_interleave4x4_height = 1;
-    const int depth_output_gemm3d       = gemm_info.depth_output_gemm3d();
+    DataType           data_type                 = a->data_type();
+    bool               reinterpret_input_as_3d   = gemm_info.reinterpret_input_as_3d();
+    const unsigned int m                         = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
+    const unsigned int n                         = b->dimension(0);
+    const unsigned int k                         = a->dimension(0);
+    int                mult_transpose1xW_width   = 1;
+    int                mult_interleave4x4_height = 1;
+    const int          depth_output_gemm3d       = gemm_info.depth_output_gemm3d();
 
     if(get_arch_from_target(gpu_target) == GPUTarget::BIFROST)
     {
@@ -214,6 +289,9 @@ Status CLGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso
     // Check if we need to reshape the matrix A and matrix B
     const bool run_interleave_transpose = is_interleaved_transposed(m, n, k, a->data_type(), reshape_b_only_on_first_run, gpu_target);
 
+    // Check if we can run the new reshaped GEMM
+    const bool is_G76_path = (gpu_target == GPUTarget::G76) && run_interleave_transpose && (data_type == DataType::F32);
+
     // if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
     if(run_interleave_transpose)
     {
@@ -227,17 +305,41 @@ Status CLGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso
         matrix_a_info = &tmp_a_info;
         matrix_b_info = &tmp_b_info;
 
-        // Validate interleave kernel
-        auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d())));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()));
+        if(is_G76_path)
+        {
+            GEMMLHSMatrixInfo lhs_info;
+            GEMMRHSMatrixInfo rhs_info;
 
-        // Validate transpose kernel
-        auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width)));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMTranspose1xWKernel::validate(b, &tmp_b_info, mult_transpose1xW_width));
+            // Pick up the GEMM configuration based on M,N and K
+            select_gemm_configuration(m, n, lhs_info, rhs_info);
+
+            auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, gemm_info.reinterpret_input_as_3d())));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMReshapeLHSMatrixKernel::validate(a, &tmp_a_info, lhs_info, gemm_info.reinterpret_input_as_3d()));
+
+            auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMReshapeRHSMatrixKernel::validate(b, &tmp_b_info, rhs_info));
+
+            // Validate matrix multiply
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixMultiplyReshapedKernel::validate(matrix_a_info, matrix_b_info, output, alpha, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1,
+                                                                                     depth_output_gemm3d, reinterpret_input_as_3d)));
+        }
+        else
+        {
+            // Validate interleave kernel
+            auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d())));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()));
+
+            // Validate transpose kernel
+            auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width)));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMTranspose1xWKernel::validate(b, &tmp_b_info, mult_transpose1xW_width));
+        }
     }
 
-    // Validate matrix multiply
-    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output, alpha, run_interleave_transpose, reshape_info, gpu_target, gemm_info.fp_mixed_precision()));
+    if(!is_G76_path)
+    {
+        // Validate matrix multiply
+        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output, alpha, run_interleave_transpose, reshape_info, gpu_target, gemm_info.fp_mixed_precision()));
+    }
 
     if(beta != 0 && c != nullptr)
     {
@@ -257,17 +359,38 @@ void CLGEMM::run()
     if(_is_interleaved_transposed)
     {
         // Run interleave kernel
-        CLScheduler::get().enqueue(_interleave_kernel, false);
+        if(_is_G76_path)
+        {
+            CLScheduler::get().enqueue(_reshape_lhs_kernel, false);
+        }
+        else
+        {
+            CLScheduler::get().enqueue(_interleave_kernel, false);
+        }
 
         if(!_reshape_b_only_on_first_run)
         {
             // Run transpose kernel
-            CLScheduler::get().enqueue(_transpose_kernel, false);
+            if(_is_G76_path)
+            {
+                CLScheduler::get().enqueue(_reshape_rhs_kernel, false);
+            }
+            else
+            {
+                CLScheduler::get().enqueue(_transpose_kernel, false);
+            }
         }
     }
 
     // Run matrix multiply kernel
-    CLScheduler::get().enqueue(_mm_kernel, !_run_addition);
+    if(_is_G76_path)
+    {
+        CLScheduler::get().enqueue(_mm_reshaped_kernel, !_run_addition);
+    }
+    else
+    {
+        CLScheduler::get().enqueue(_mm_kernel, !_run_addition);
+    }
 
     // Run matrix addition kernel
     if(_run_addition)
@@ -286,7 +409,14 @@ void CLGEMM::prepare()
         {
             // Run transpose kernel and mark original weights tensor as unused
             _tmp_b.allocator()->allocate();
-            CLScheduler::get().enqueue(_transpose_kernel, false);
+            if(_is_G76_path)
+            {
+                CLScheduler::get().enqueue(_reshape_rhs_kernel, false);
+            }
+            else
+            {
+                CLScheduler::get().enqueue(_transpose_kernel, false);
+            }
             _original_b->mark_as_unused();
         }
         CLScheduler::get().queue().finish();
diff --git a/tests/framework/Macros.h b/tests/framework/Macros.h
index deca1ef51a..591b80e9d8 100644
--- a/tests/framework/Macros.h
+++ b/tests/framework/Macros.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -49,8 +49,8 @@
 
 #define CONCAT(ARG0, ARG1) ARG0##ARG1
 
-#define VARIADIC_SIZE_IMPL(e0, e1, e2, e3, e4, e5, e6, e7, e8, e9, size, ...) size
-#define VARIADIC_SIZE(...) VARIADIC_SIZE_IMPL(__VA_ARGS__, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)
+#define VARIADIC_SIZE_IMPL(e0, e1, e2, e3, e4, e5, e6, e7, e8, e9, e10, size, ...) size
+#define VARIADIC_SIZE(...) VARIADIC_SIZE_IMPL(__VA_ARGS__, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)
 
 #define JOIN_PARAM1(OP, param) OP(0, param)
 #define JOIN_PARAM2(OP, param, ...) \
@@ -80,6 +80,9 @@
 #define JOIN_PARAM10(OP, param, ...) \
     OP(9, param)                     \
     , JOIN_PARAM9(OP, __VA_ARGS__)
+#define JOIN_PARAM11(OP, param, ...) \
+    OP(10, param)                    \
+    , JOIN_PARAM10(OP, __VA_ARGS__)
 #define JOIN_PARAM(OP, NUM, ...) \
     CONCAT(JOIN_PARAM, NUM)      \
     (OP, __VA_ARGS__)
@@ -264,4 +267,4 @@
 //
 // TEST CASE MACROS END
 //
-#endif /* ARM_COMPUTE_TEST_FRAMEWORK_MACROS */
+#endif /* ARM_COMPUTE_TEST_FRAMEWORK_MACROS */
\ No newline at end of file
diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
new file mode 100644
index 0000000000..e2fa194765
--- /dev/null
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
@@ -0,0 +1,224 @@
+/*
+ * 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/CLGEMMMatrixMultiplyReshapedKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.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
+{
+namespace
+{
+// *INDENT-OFF*
+// clang-format off
+RelativeTolerance<float> tolerance_f32(0.001f);
+constexpr float          abs_tolerance_f32(0.0001f);
+
+/** M values to test */
+const auto m_values = framework::dataset::make("M", 37);
+
+/** 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", 43);
+
+/** 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, 5, 6});
+
+/** N0 values to test - Precommit */
+const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4, 8 });
+
+/** K0 values to test - Precommit */
+const auto k0_values_precommit = framework::dataset::make("K0", { 4, 8 });
+
+/** V0 values to test - Precommit */
+const auto v0_values_precommit = framework::dataset::make("V0", 1, 3);
+
+/** 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, 4, 8, 16 });
+
+/** K0 values to test - Nightly */
+const auto k0_values_nightly = framework::dataset::make("K0", { 4, 8, 16 });
+
+/** V0 values to test - Nightly */
+const auto v0_values_nightly = framework::dataset::make("V0", 1, 4);
+
+/** H0 values to test - Nightly */
+const auto h0_values_nightly = framework::dataset::make("H0", 1, 4);
+
+/** Interleave values to test with LHS matrix */
+const auto i_values_lhs = framework::dataset::make("interleave_lhs", { true, false });
+
+/** Interleave values to test with RHS matrix */
+const auto i_values_rhs = framework::dataset::make("interleave_rhs", { true, false });
+
+} // namespace
+
+using namespace arm_compute::misc::shape_calculator;
+
+// Create function for CLGEMMReshapeLHSMatrixKernel
+using CLGEMMReshapeLHSMatrix = CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder<CLGEMMReshapeLHSMatrixKernel, 16>;
+
+// Create function for CLGEMMReshapeRHSMatrixKernel
+using CLGEMMReshapeRHSMatrix = CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder<CLGEMMReshapeRHSMatrixKernel, 16>;
+
+// Create function for CLGEMMMatrixMultiplyReshapedKernel
+using CLGEMMMatrixMultiplyReshaped = CLSynthetizeFunction<CLGEMMMatrixMultiplyReshapedKernel>;
+
+// Fixture for CLGEMMMatrixMultiplyReshaped
+using CLGEMMMatrixMultiplyReshapedFixture = GEMMMatrixMultiplyReshapedValidationFixture<CLTensor, CLAccessor, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>;
+
+TEST_SUITE(CL)
+TEST_SUITE(GEMMMatrixMultiplyReshaped)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(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),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs),
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs)
+{
+    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;
+    lhs_info.v0         = v0_value;
+    lhs_info.interleave = i_value_lhs;
+    lhs_info.transpose  = false;
+
+    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 lhs_shape_reshaped = compute_lhs_reshaped_shape(TensorInfo(lhs_shape, 1, DataType::F32),
+                                                                      lhs_info,
+                                                                      false);
+
+    const TensorShape rhs_shape(N, K, b_value);
+    const TensorShape rhs_shape_reshaped = compute_rhs_reshaped_shape(TensorInfo(rhs_shape, 1, DataType::F32),
+                                                                      rhs_info);
+
+    const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape_reshaped, 1, DataType::F32),
+                                                   TensorInfo(rhs_shape_reshaped, 1, DataType::F32),
+                                                   gemm_info);
+
+    // Create tensors
+    CLTensor lhs_reshaped = create_tensor<CLTensor>(lhs_shape_reshaped, DataType::F32);
+    CLTensor rhs_reshaped = create_tensor<CLTensor>(rhs_shape_reshaped, DataType::F32);
+    CLTensor dst          = create_tensor<CLTensor>(dst_shape, DataType::F32);
+
+    ARM_COMPUTE_EXPECT(lhs_reshaped.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
+    CLGEMMMatrixMultiplyReshaped gemm;
+    gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, 1.0f, lhs_info, rhs_info, gemm_info);
+}
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture, framework::DatasetMode::ALL,
+                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),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture, framework::DatasetMode::NIGHTLY,
+                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),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+}
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // GEMMMatrixMulipltyReshaped
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/CL/Im2Col.cpp b/tests/validation/CL/Im2Col.cpp
index ebf2331e5e..432b3b239a 100644
--- a/tests/validation/CL/Im2Col.cpp
+++ b/tests/validation/CL/Im2Col.cpp
@@ -49,6 +49,7 @@ const auto conv_filter_sizes = framework::dataset::make("KernelDims", { Size2D(3
                                                                         Size2D(1U, 3U),
                                                                         Size2D(5U, 3U),
                                                                         Size2D(1U, 1U),
+                                                                        Size2D(9U, 9U),
                                                                         Size2D(11U, 11U)} );
 const auto padstrides        = framework::dataset::make("PadStride", { PadStrideInfo(1U, 1U, 0U, 0U),
                                                                        PadStrideInfo(1U, 1U, 1U, 1U),
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index 0083abffb5..ce2b177ce9 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -151,6 +151,117 @@ protected:
     SimpleTensor<T> _reference{};
 };
 
+template <typename TensorType, typename AccessorType, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMMatrixMultiplyReshapedValidationFixture : 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 v0, unsigned int h0, bool interleave_lhs,
+               bool interleave_rhs)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0         = m0;
+        lhs_info.k0         = k0;
+        lhs_info.v0         = v0;
+        lhs_info.interleave = interleave_lhs;
+        lhs_info.transpose  = false;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0         = n0;
+        rhs_info.k0         = k0;
+        rhs_info.h0         = h0;
+        rhs_info.interleave = interleave_rhs;
+        rhs_info.transpose  = true;
+
+        // 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);
+        _reference = compute_reference(lhs_shape, rhs_shape);
+    }
+
+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)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, DataType::F32, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, DataType::F32, 1);
+        TensorType lhs_reshaped;
+        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
+        ReshapeLHSFunctionType reshape_lhs;
+        ReshapeRHSFunctionType reshape_rhs;
+        GEMMFunctionType       gemm;
+        reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info);
+        reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
+        gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, 1.0f, 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();
+        lhs_reshaped.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(!lhs_reshaped.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_lhs.run();
+        reshape_rhs.run();
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<float> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape[0]          = rhs_shape[0];
+        dst_shape[1]          = lhs_shape[1];
+
+        // Create reference
+        SimpleTensor<float> lhs{ lhs_shape, DataType::F32, 1 };
+        SimpleTensor<float> rhs{ rhs_shape, DataType::F32, 1 };
+        SimpleTensor<float> c{ dst_shape, DataType::F32, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+        fill(c, 2);
+
+        return reference::gemm<float>(lhs, rhs, c, 1.0f, 0.0f);
+    }
+
+    TensorType          _target{};
+    SimpleTensor<float> _reference{};
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute