COMPMID-1698: Implementing CLGEMMLowpMatrixMultiplyReshapedKernel

Change-Id: Ia4db21b394a0b9235393202ce3c00b11cceb94ea
Reviewed-on: https://review.mlplatform.org/568
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

12 files changed, 1580 insertions, 153 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index e68769b6ae..07e214be3f 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -69,6 +69,7 @@
 #include "arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h
new file mode 100644
index 0000000000..1cf7236446
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h
@@ -0,0 +1,86 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYRESHAPEDKERNEL_H__
+#define __ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYRESHAPEDKERNEL_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 CLGEMMLowpMatrixMultiplyReshapedKernel : public ICLKernel
+{
+public:
+    /** Default Constructor */
+    CLGEMMLowpMatrixMultiplyReshapedKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMLowpMatrixMultiplyReshapedKernel(const CLGEMMLowpMatrixMultiplyReshapedKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMLowpMatrixMultiplyReshapedKernel &operator=(const CLGEMMLowpMatrixMultiplyReshapedKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLGEMMLowpMatrixMultiplyReshapedKernel(CLGEMMLowpMatrixMultiplyReshapedKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLGEMMLowpMatrixMultiplyReshapedKernel &operator=(CLGEMMLowpMatrixMultiplyReshapedKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  input0    Input tensor containing the LHS reshaped matrix. Data type supported: QASYMM8
+     * @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]  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, 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 CLGEMMLowpMatrixMultiplyReshapedKernel
+     *
+     * @param[in] input0    Input tensor info containing the LHS reshaped matrix. Data type supported: QASYMM8
+     * @param[in] input1    Input tensor info containing the RHS reshaped matrix. Data type supported: same as @p input0
+     * @param[in] output    Output tensor info. Data type supported: same as @p input0
+     * @param[in] 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, 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;
+    unsigned int     _k;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYRESHAPEDKERNEL_H__*/
\ No newline at end of file
diff --git a/arm_compute/runtime/CL/functions/CLGEMM.h b/arm_compute/runtime/CL/functions/CLGEMM.h
index 3ec07cf5f9..0bad446551 100644
--- a/arm_compute/runtime/CL/functions/CLGEMM.h
+++ b/arm_compute/runtime/CL/functions/CLGEMM.h
@@ -115,7 +115,7 @@ private:
     bool                               _run_addition;
     bool                               _reshape_b_only_on_first_run;
     bool                               _is_prepared;
-    bool                               _is_new_gemm_reshaped; // Removed when COMPMID-1892 is completed
+    bool                               _is_new_gemm_reshaped; // Remove when COMPMID-1892 is completed
 };
 } // namespace arm_compute
 
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
index 72d91070f8..4345ff267b 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
@@ -25,6 +25,7 @@
 #define __ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYCORE_H__
 
 #include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h"
@@ -43,7 +44,8 @@ class ICLTensor;
  *
  *  -# @ref CLGEMMReshapeLHSMatrixKernel  (if the output tensor is a matrix)
  *  -# @ref CLGEMMReshapeRHSMatrixKernel  (if the output tensor is a matrix)
- *  -# @ref CLGEMMLowpMatrixMultiplyKernel
+ *  -# @ref CLGEMMLowpMatrixMultiplyKernel (if the input matrix is a vector or for Midgard architectures)
+ *  -# @ref CLGEMMLowpMatrixMultiplyReshapedKernel (if the input matrix is not a vector and if the GPU architecture is not Midgard)
  *  -# @ref CLGEMMLowpMatrixAReductionKernel (if the offset of matrix B is not 0)
  *  -# @ref CLGEMMLowpMatrixBReductionKernel (if the offset of matrix A is not 0)
  *  -# @ref CLGEMMLowpOffsetContributionKernel (if gemm_info.gemmlowp_output_stage == NONE)
@@ -101,6 +103,7 @@ public:
 private:
     CLMemoryGroup                                 _memory_group;
     CLGEMMLowpMatrixMultiplyKernel                _mm_kernel;
+    CLGEMMLowpMatrixMultiplyReshapedKernel        _mm_reshaped_kernel;
     CLGEMMReshapeLHSMatrixKernel                  _mtx_a_reshape_kernel;
     CLGEMMReshapeRHSMatrixKernel                  _mtx_b_reshape_kernel;
     CLGEMMLowpMatrixAReductionKernel              _mtx_a_reduction_kernel;
@@ -115,10 +118,10 @@ private:
     const ICLTensor                              *_original_b;
     int32_t                                       _a_offset;
     int32_t                                       _b_offset;
-    bool                                          _is_interleaved_transposed;
+    bool                                          _is_gemm_reshaped;
     bool                                          _reshape_b_only_on_first_run;
     bool                                          _is_prepared;
     bool                                          _fuse_output_stage;
 };
 }
-#endif /*__ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYCORE_H__ */
+#endif /*__ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYCORE_H__ */
\ No newline at end of file
diff --git a/arm_compute/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.h b/arm_compute/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.h
index b452c53c39..c452e159cf 100644
--- a/arm_compute/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.h
+++ b/arm_compute/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.h
@@ -40,6 +40,8 @@ public:
 private:
     std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
     std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+    std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+    std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
 };
 } // namespace cl_gemm
 } // namespace arm_compute
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 905a34a509..4635d11a3a 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -297,6 +297,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemmlowp_mm_interleaved_transposed_bifrost", "gemmlowp.cl" },
     { "gemmlowp_mm_interleaved_transposed_bifrost_dot8", "gemmlowp.cl" },
     { "gemmlowp_mm_interleaved_transposed_midgard", "gemmlowp.cl" },
+    { "gemmlowp_mm_reshaped_lhs_nt_rhs_t", "gemmlowp.cl" },
+    { "gemmlowp_mm_reshaped_lhs_nt_rhs_t_dot8", "gemmlowp.cl" },
     { "gemmlowp_offset_contribution", "gemmlowp.cl" },
     { "gemmlowp_offset_contribution_quantize_down", "gemmlowp.cl" },
     { "gemmlowp_offset_contribution_quantize_down_fixedpoint", "gemmlowp.cl" },
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index 8c1fa548e4..277338bf08 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,6 +23,7 @@
  */
 #include "helpers.h"
 #include "helpers_asymm.h"
+#include "repeat.h"
 
 #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
 #if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
@@ -1943,6 +1944,574 @@ __kernel void gemmlowp_mm_bifrost_dot8(IMAGE_DECLARATION(src0),
 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
 #endif // defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_Y) && defined(COLS_A)
 
+#if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0)
+
+#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
+
+#if K0 == 2
+#define ARM_DOT_K0(a, b, c)                                         \
+    ({                                                              \
+        ARM_DOT((uchar4)(a, (uchar2)0), (uchar4)(b, (uchar2)0), c); \
+    })
+#elif K0 == 3 // K0 == 3
+#define ARM_DOT_K0(a, b, c)                                       \
+    ({                                                            \
+        ARM_DOT((uchar4)(a, (uchar)0), (uchar4)(b, (uchar)0), c); \
+    })
+#elif K0 == 4 // 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
+
+#else // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
+
+#if K0 == 2
+#define ARM_DOT_K0(a, b, c)     \
+    ({                          \
+        c += (uint)a.s0 * b.s0; \
+        c += (uint)a.s1 * b.s1; \
+    })
+#elif K0 == 3 // K0 == 3
+#define ARM_DOT_K0(a, b, c)     \
+    ({                          \
+        c += (uint)a.s0 * b.s0; \
+        c += (uint)a.s1 * b.s1; \
+        c += (uint)a.s2 * b.s2; \
+    })
+#elif K0 == 4 // K0 == 4
+#define ARM_DOT_K0(a, b, c)     \
+    ({                          \
+        c += (uint)a.s0 * b.s0; \
+        c += (uint)a.s1 * b.s1; \
+        c += (uint)a.s2 * b.s2; \
+        c += (uint)a.s3 * b.s3; \
+    })
+#elif K0 == 8 // K0 == 8
+#define ARM_DOT_K0(a, b, c)     \
+    ({                          \
+        c += (uint)a.s0 * b.s0; \
+        c += (uint)a.s1 * b.s1; \
+        c += (uint)a.s2 * b.s2; \
+        c += (uint)a.s3 * b.s3; \
+        c += (uint)a.s4 * b.s4; \
+        c += (uint)a.s5 * b.s5; \
+        c += (uint)a.s6 * b.s6; \
+        c += (uint)a.s7 * b.s7; \
+    })
+#elif K0 == 16 // K0 == 16
+#define ARM_DOT_K0(a, b, c)     \
+    ({                          \
+        c += (uint)a.s0 * b.s0; \
+        c += (uint)a.s1 * b.s1; \
+        c += (uint)a.s2 * b.s2; \
+        c += (uint)a.s3 * b.s3; \
+        c += (uint)a.s4 * b.s4; \
+        c += (uint)a.s5 * b.s5; \
+        c += (uint)a.s6 * b.s6; \
+        c += (uint)a.s7 * b.s7; \
+        c += (uint)a.s8 * b.s8; \
+        c += (uint)a.s9 * b.s9; \
+        c += (uint)a.sA * b.sA; \
+        c += (uint)a.sB * b.sB; \
+        c += (uint)a.sC * b.sC; \
+        c += (uint)a.sD * b.sD; \
+        c += (uint)a.sE * b.sE; \
+        c += (uint)a.sF * b.sF; \
+    })
+#else // K0 not supported
+#error "K0 value not supported"
+#endif // K0
+
+#endif //defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
+
+#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 == 3 // N0 == 3
+#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)); \
+    })
+#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 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, 3, 4, 8, 16
+ *  - K0 = 2, 3, 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: QASYMM8
+ * @param[in]  lhs_stride_x                      Stride of the LHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  lhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  lhs_stride_y                      Stride of the LHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  lhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
+ * @param[in]  rhs_ptr                           Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  rhs_stride_x                      Stride of the RHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  rhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  rhs_stride_y                      Stride of the RHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  rhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: same as @p lhs_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
+ * @param[in]  k                                 Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ * @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 gemmlowp_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs),
+                                                IMAGE_DECLARATION(rhs),
+                                                IMAGE_DECLARATION(dst),
+                                                uint k,
+                                                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 + (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 + (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
+    REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(uint, N0), c, 0); //VEC_DATA_TYPE(uint, N0)    c0=0,c1=0,c2=0,... c(M0-1)=0;
+
+    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(uchar, K0)
+        a0 = VLOAD(K0)(0, lhs_addr + 0 * LHS_STEP_X);
+#if M0 > 1
+        VEC_DATA_TYPE(uchar, K0)
+        a1 = VLOAD(K0)(0, lhs_addr + 1 * LHS_STEP_X);
+#endif // M0 > 1
+#if M0 > 2
+        VEC_DATA_TYPE(uchar, K0)
+        a2 = VLOAD(K0)(0, lhs_addr + 2 * LHS_STEP_X);
+#endif // M0 > 2
+#if M0 > 3
+        VEC_DATA_TYPE(uchar, K0)
+        a3 = VLOAD(K0)(0, lhs_addr + 3 * LHS_STEP_X);
+#endif // M0 > 3
+#if M0 > 4
+        VEC_DATA_TYPE(uchar, K0)
+        a4 = VLOAD(K0)(0, lhs_addr + 4 * LHS_STEP_X);
+#endif // M0 > 4
+#if M0 > 5
+        VEC_DATA_TYPE(uchar, K0)
+        a5 = VLOAD(K0)(0, lhs_addr + 5 * LHS_STEP_X);
+#endif // M0 > 5
+#if M0 > 6
+        VEC_DATA_TYPE(uchar, K0)
+        a6 = VLOAD(K0)(0, lhs_addr + 6 * LHS_STEP_X);
+#endif // M0 > 6
+#if M0 > 7
+        VEC_DATA_TYPE(uchar, K0)
+        a7 = VLOAD(K0)(0, lhs_addr + 7 * LHS_STEP_X);
+#endif // M0 > 7
+
+        // Load values from RHS matrix
+        VEC_DATA_TYPE(uchar, K0)
+        b0 = VLOAD(K0)(0, rhs_addr + 0 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        b1 = VLOAD(K0)(0, rhs_addr + 1 * RHS_STEP_X);
+#if N0 > 2
+        VEC_DATA_TYPE(uchar, K0)
+        b2 = VLOAD(K0)(0, rhs_addr + 2 * RHS_STEP_X);
+#endif // N0 > 2
+#if N0 > 3
+        VEC_DATA_TYPE(uchar, K0)
+        b3 = VLOAD(K0)(0, rhs_addr + 3 * RHS_STEP_X);
+#endif // N0 > 3
+#if N0 > 4
+        VEC_DATA_TYPE(uchar, K0)
+        b4 = VLOAD(K0)(0, rhs_addr + 4 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        b5 = VLOAD(K0)(0, rhs_addr + 5 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        b6 = VLOAD(K0)(0, rhs_addr + 6 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        b7 = VLOAD(K0)(0, rhs_addr + 7 * RHS_STEP_X);
+#endif // N0 > 4
+#if N0 > 8
+        VEC_DATA_TYPE(uchar, K0)
+        b8 = VLOAD(K0)(0, rhs_addr + 8 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        b9 = VLOAD(K0)(0, rhs_addr + 9 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        bA = VLOAD(K0)(0, rhs_addr + 10 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        bB = VLOAD(K0)(0, rhs_addr + 11 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        bC = VLOAD(K0)(0, rhs_addr + 12 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        bD = VLOAD(K0)(0, rhs_addr + 13 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        bE = VLOAD(K0)(0, rhs_addr + 14 * RHS_STEP_X);
+        VEC_DATA_TYPE(uchar, K0)
+        bF = VLOAD(K0)(0, rhs_addr + 15 * RHS_STEP_X);
+#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);
+        rhs_addr += (N0 * RHS_STEP_X * RHS_STEP_LOOP);
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(int)) + (get_global_id(1) * (uint)M0 * dst_stride_y);
+
+    REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
+
+#if defined(REINTERPRET_OUTPUT_AS_3D)
+    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
+    // in order to take into account the presence of possible cross plane paddings
+    //
+    //  |                  |
+    //  |      plane0      |
+    //  |                  |
+    //  |__________________|
+    //  |******************|
+    //  |  cross_plane_pad |
+    //  |******************|
+    //  |                  |
+    //  |      plane1      |
+    //  |                  |
+    //  |__________________|
+
+    // The plane (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_y);
+#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_y);
+#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_y);
+#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_y);
+#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_y);
+#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_y);
+#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_y);
+#endif // M0 > 6
+#if M0 > 7
+    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_y);
+#endif // M0 > 7
+
+    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
+    // multiply dst_stride_z by DEPTH_GEMM3D
+    dst_addr += 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)
+
+    // Store output block
+    VSTORE(N0)
+    (CONVERT_SAT(c0, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 0 * dst_stride_y + zout0));
+#if M0 > 1
+    VSTORE(N0)
+    (CONVERT_SAT(c1, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 1 * dst_stride_y + zout1));
+#endif // M0 > 1
+#if M0 > 2
+    VSTORE(N0)
+    (CONVERT_SAT(c2, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 2 * dst_stride_y + zout2));
+#endif // M0 > 2
+#if M0 > 3
+    VSTORE(N0)
+    (CONVERT_SAT(c3, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 3 * dst_stride_y + zout3));
+#endif // M0 > 3
+#if M0 > 4
+    VSTORE(N0)
+    (CONVERT_SAT(c4, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 4 * dst_stride_y + zout4));
+#endif // M0 > 4
+#if M0 > 5
+    VSTORE(N0)
+    (CONVERT_SAT(c5, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 5 * dst_stride_y + zout5));
+#endif // M0 > 5
+#if M0 > 6
+    VSTORE(N0)
+    (CONVERT_SAT(c6, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 6 * dst_stride_y + zout6));
+#endif // M0 > 6
+#if M0 > 7
+    VSTORE(N0)
+    (CONVERT_SAT(c7, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(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
+}
+
+#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices unsing the dot8 instruction.
+ *  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 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, 3, 4, 8, 16
+ *  - K0 = 2, 3, 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: QASYMM8
+ * @param[in]  lhs_stride_x                      Stride of the LHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  lhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  lhs_stride_y                      Stride of the LHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  lhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
+ * @param[in]  rhs_ptr                           Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  rhs_stride_x                      Stride of the RHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  rhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  rhs_stride_y                      Stride of the RHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  rhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: same as @p lhs_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
+ * @param[in]  k                                 Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ * @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 gemmlowp_mm_reshaped_lhs_nt_rhs_t_dot8(IMAGE_DECLARATION(lhs),
+                                                     IMAGE_DECLARATION(rhs),
+                                                     IMAGE_DECLARATION(dst),
+                                                     uint k,
+                                                     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
+                                                    )
+{
+    // Note: ARM_DOT_K0XN0 is generated with the dot8 instruction
+    gemmlowp_mm_reshaped_lhs_nt_rhs_t(lhs_ptr,
+                                      lhs_stride_x,
+                                      lhs_step_x,
+                                      lhs_stride_y,
+                                      lhs_step_y,
+                                      lhs_offset_first_element_in_bytes,
+                                      rhs_ptr,
+                                      rhs_stride_x,
+                                      rhs_step_x,
+                                      rhs_stride_y,
+                                      rhs_step_y,
+                                      rhs_offset_first_element_in_bytes,
+                                      dst_ptr,
+                                      dst_stride_x,
+                                      dst_step_x,
+                                      dst_stride_y,
+                                      dst_step_y,
+                                      dst_offset_first_element_in_bytes,
+                                      k,
+                                      lhs_stride_z,
+                                      rhs_stride_z,
+                                      dst_stride_z
+#if defined(REINTERPRET_OUTPUT_AS_3D)
+                                      ,
+                                      dst_cross_plane_pad
+#endif // REINTERPRET_OUTPUT_AS_3D
+                                     );
+}
+#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
+#endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(K)
+
 #if defined(COLS_A)
 /** OpenCL kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A.
  *
diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.cpp
new file mode 100644
index 0000000000..e9be1a6dfc
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.cpp
@@ -0,0 +1,308 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.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, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                          const GEMMReshapeInfo &gemm_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);
+    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_MSG(((lhs_info.k0 & (lhs_info.k0 - 1)) && lhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");
+
+    const int m = gemm_info.m();
+    const int n = gemm_info.n();
+    const int k = gemm_info.k();
+
+    TensorShape tensor_shape0{ input0->tensor_shape() };
+    tensor_shape0.set(0, k);
+    tensor_shape0.set(1, m);
+
+    TensorShape tensor_shape1{ input1->tensor_shape() };
+    tensor_shape1.set(0, n);
+    tensor_shape1.set(1, k);
+
+    const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
+    const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
+
+    const TensorInfo tensor_info_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_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
+    }
+
+    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)).set_data_type(DataType::S32));
+
+    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
+
+CLGEMMLowpMatrixMultiplyReshapedKernel::CLGEMMLowpMatrixMultiplyReshapedKernel()
+    : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1)
+{
+}
+
+void CLGEMMLowpMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, 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(), lhs_info, rhs_info, gemm_info));
+
+    _input0                   = input0;
+    _input1                   = input1;
+    _output                   = output;
+    _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
+    _k                        = gemm_info.k();
+
+    // 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_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("-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("gemmlowp_mm_reshaped_");
+    kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";
+    kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt";
+    kernel_name += dot8_supported(CLKernelLibrary::get().get_device()) ? "_dot8" : "";
+
+    // 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 += 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 CLGEMMLowpMatrixMultiplyReshapedKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, 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, 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 CLGEMMLowpMatrixMultiplyReshapedKernel::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() + 4;
+        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>(_k));
+        _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/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index 4b72878b5f..2a01db7824 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -31,43 +31,25 @@
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
+#include "arm_compute/runtime/CL/gemm_reshaped/CLGEMMReshapedConfiguration.h"
 
 namespace arm_compute
 {
 using namespace arm_compute::misc::shape_calculator;
+using namespace arm_compute::cl_gemm;
 
 namespace
 {
-inline bool is_interleaved_transposed(int m, int n, int k, bool reshape_b_only_on_first_run, GPUTarget gpu_target)
+inline bool is_gemm_reshaped(unsigned int m, bool reshape_b_only_on_first_run, GPUTarget gpu_target)
 {
-    bool flag = true;
-
-    if(gpu_target_is_in(gpu_target,
-                        GPUTarget::G71, GPUTarget::G72,
-                        GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT))
-    {
-        // COMPMID-852
-        if(k > 256 && m > 4 && reshape_b_only_on_first_run)
-        {
-            flag = ((0.72f + n * 0.10766f) < (n * 0.1284f));
-        }
-        else
-        {
-            flag = false;
-        }
-    }
-    else
-    {
-        flag = m > 1;
-    }
-
-    return flag;
+    return (get_arch_from_target(gpu_target) != GPUTarget::MIDGARD) && (m > 1) && (reshape_b_only_on_first_run);
 }
 } // namespace
 
 CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)),
       _mm_kernel(),
+      _mm_reshaped_kernel(),
       _mtx_a_reshape_kernel(),
       _mtx_b_reshape_kernel(),
       _mtx_a_reduction_kernel(),
@@ -82,7 +64,7 @@ CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemo
       _original_b(nullptr),
       _a_offset(0),
       _b_offset(0),
-      _is_interleaved_transposed(true),
+      _is_gemm_reshaped(true),
       _reshape_b_only_on_first_run(false),
       _is_prepared(false),
       _fuse_output_stage(false)
@@ -115,29 +97,17 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, 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 bool    unroll_block              = dot8_supported(CLKernelLibrary::get().get_device());
-    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();
-    constexpr int mult_transpose1xW_width   = 1;
-    constexpr int mult_interleave4x4_height = 1;
-    rhs_info.n0                             = 16 / b->info()->element_size();
-    rhs_info.k0                             = 1;
-    rhs_info.h0                             = mult_transpose1xW_width;
-    rhs_info.interleave                     = false;
-    rhs_info.transpose                      = false;
-    lhs_info.m0                             = 4;
-    lhs_info.k0                             = 4;
-    lhs_info.v0                             = mult_interleave4x4_height;
-    lhs_info.interleave                     = true;
-    lhs_info.transpose                      = !unroll_block;
+    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 unsigned int batch_size              = reinterpret_input_as_3d ? a->info()->dimension(3) : a->info()->dimension(2);
+    const int          depth_output_gemm3d     = gemm_info.depth_output_gemm3d();
 
     // Check if we need to reshape the matrix A and matrix B
-    _is_interleaved_transposed = is_interleaved_transposed(m, n, k, _reshape_b_only_on_first_run, gpu_target);
+    _is_gemm_reshaped = is_gemm_reshaped(m, _reshape_b_only_on_first_run, gpu_target);
 
-    if(_is_interleaved_transposed)
+    if(_is_gemm_reshaped)
     {
         // if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
         reinterpret_input_as_3d = false;
@@ -151,6 +121,9 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
             _memory_group.manage(&_tmp_b);
         }
 
+        // Pick up the GEMM configuration
+        std::tie(lhs_info, rhs_info) = CLGEMMReshapedConfigurationFactory::create()->configure(m, n, k, batch_size, DataType::QASYMM8);
+
         // Configure interleave kernel
         _mtx_a_reshape_kernel.configure(a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d());
 
@@ -190,10 +163,16 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
 
         _memory_group.manage(&_mm_result_s32);
 
-        // Configure matrix multiply kernel
-        _mm_kernel.configure(matrix_a, matrix_b, &_mm_result_s32, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k,
-                                                                                                              mult_transpose1xW_width, mult_interleave4x4_height,
-                                                                                                              depth_output_gemm3d, reinterpret_input_as_3d));
+        if(_is_gemm_reshaped)
+        {
+            // Configure and tune matrix multiply kernel
+            _mm_reshaped_kernel.configure(matrix_a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+        }
+        else
+        {
+            // Configure matrix multiply kernel
+            _mm_kernel.configure(matrix_a, matrix_b, &_mm_result_s32, false, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+        }
 
         // Configure offset contribution kernel
         _offset_contribution_output_stage_kernel.configure(&_mm_result_s32, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c, output, a->info()->dimension(0),
@@ -203,17 +182,23 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
     }
     else
     {
-        // Configure matrix multiply kernel
-        _mm_kernel.configure(matrix_a, matrix_b, output, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k,
-                                                                                                     mult_transpose1xW_width, mult_interleave4x4_height,
-                                                                                                     depth_output_gemm3d, reinterpret_input_as_3d));
+        if(_is_gemm_reshaped)
+        {
+            // Configure and tune matrix multiply kernel
+            _mm_reshaped_kernel.configure(matrix_a, matrix_b, output, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+        }
+        else
+        {
+            // Configure matrix multiply kernel
+            _mm_kernel.configure(matrix_a, matrix_b, output, false, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+        }
 
         // Configure offset contribution kernel
         _offset_contribution_kernel.configure(output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c, a->info()->dimension(0), _a_offset, _b_offset);
     }
 
     // Allocate tensors
-    if(_is_interleaved_transposed)
+    if(_is_gemm_reshaped)
     {
         _tmp_a.allocator()->allocate();
         if(!_reshape_b_only_on_first_run)
@@ -251,26 +236,14 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
     GEMMRHSMatrixInfo rhs_info;
     GEMMLHSMatrixInfo lhs_info;
 
-    bool          reinterpret_input_as_3d   = gemm_info.reinterpret_input_as_3d();
-    const bool    unroll_block              = dot8_supported(CLKernelLibrary::get().get_device());
-    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);
-    constexpr int mult_transpose1xW_width   = 1;
-    constexpr int mult_interleave4x4_height = 1;
-    const int     depth_output_gemm3d       = gemm_info.depth_output_gemm3d();
-    rhs_info.n0                             = 16 / b->element_size();
-    rhs_info.k0                             = 1;
-    rhs_info.h0                             = mult_transpose1xW_width;
-    rhs_info.interleave                     = false;
-    rhs_info.transpose                      = false;
-    lhs_info.m0                             = 4;
-    lhs_info.k0                             = 4;
-    lhs_info.v0                             = mult_interleave4x4_height;
-    lhs_info.interleave                     = true;
-    lhs_info.transpose                      = !unroll_block;
-
-    bool reshape_matrices = is_interleaved_transposed(m, n, k, gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
+    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);
+    const unsigned int batch_size              = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2);
+    const int          depth_output_gemm3d     = gemm_info.depth_output_gemm3d();
+
+    bool reshape_matrices = is_gemm_reshaped(m, gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
 
     // if reshape_matrices is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
     if(reshape_matrices)
@@ -278,13 +251,16 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
         reinterpret_input_as_3d = false;
     }
 
-    const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d, reinterpret_input_as_3d);
+    const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d);
 
     if(reshape_matrices)
     {
         matrix_a_info = &tmp_a_info;
         matrix_b_info = &tmp_b_info;
 
+        // Pick up the GEMM configuration
+        std::tie(lhs_info, rhs_info) = CLGEMMReshapedConfigurationFactory::create()->configure(m, n, k, batch_size, DataType::QASYMM8);
+
         // Validate interleave kernel
         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()));
@@ -319,12 +295,22 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
     {
         TensorInfo mm_result_s32_info{};
 
-        // Output tensor auto inizialitation if not yet initialized
-        auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, reshape_matrices, reshape_info)).set_data_type(DataType::S32));
+        if(reshape_matrices)
+        {
+            // Output tensor auto inizialitation if not yet initialized
+            auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, reshape_info)).set_data_type(DataType::S32));
 
-        // Validate matrix multiply
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, reshape_matrices, reshape_info));
+            // Validate matrix multiply
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info));
+        }
+        else
+        {
+            // Output tensor auto inizialitation if not yet initialized
+            auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, false, reshape_info)).set_data_type(DataType::S32));
 
+            // Validate matrix multiply
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, false, reshape_info));
+        }
         // Validate offset contribution kernel
         ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOffsetContributionOutputStageKernel::validate(&mm_result_s32_info,
                                                                                             a_offset == 0 ? nullptr : &info_vector_sum_col,
@@ -336,9 +322,16 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
     }
     else
     {
-        // Validate matrix multiply
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output, reshape_matrices, reshape_info));
-
+        if(reshape_matrices)
+        {
+            // Validate matrix multiply
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedKernel::validate(matrix_a_info, matrix_b_info, output, lhs_info, rhs_info, reshape_info));
+        }
+        else
+        {
+            // Validate matrix multiply
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output, false, reshape_info));
+        }
         // Validate offset contribution kernel
         ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOffsetContributionKernel::validate(output,
                                                                                  a_offset == 0 ? nullptr : &info_vector_sum_col,
@@ -356,7 +349,7 @@ void CLGEMMLowpMatrixMultiplyCore::run()
 
     _memory_group.acquire();
 
-    if(_is_interleaved_transposed)
+    if(_is_gemm_reshaped)
     {
         // Run reshape matrix A
         CLScheduler::get().enqueue(_mtx_a_reshape_kernel, false);
@@ -375,7 +368,14 @@ void CLGEMMLowpMatrixMultiplyCore::run()
     }
 
     // Run matrix multiply
-    CLScheduler::get().enqueue(_mm_kernel, false);
+    if(_is_gemm_reshaped)
+    {
+        CLScheduler::get().enqueue(_mm_reshaped_kernel, false);
+    }
+    else
+    {
+        CLScheduler::get().enqueue(_mm_kernel, false);
+    }
 
     // Run matrix A reduction kernel only if _b_offset is not equal to 0
     if(_b_offset != 0)
@@ -401,7 +401,7 @@ void CLGEMMLowpMatrixMultiplyCore::prepare()
 {
     if(!_is_prepared)
     {
-        if(_is_interleaved_transposed && _reshape_b_only_on_first_run)
+        if(_is_gemm_reshaped && _reshape_b_only_on_first_run)
         {
             ARM_COMPUTE_ERROR_ON(!_original_b->is_used());
 
diff --git a/src/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.cpp b/src/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.cpp
index 079a52e61c..cd97849712 100644
--- a/src/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.cpp
+++ b/src/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.cpp
@@ -32,18 +32,62 @@ namespace arm_compute
 {
 namespace cl_gemm
 {
+namespace
+{
+std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_gemm_reshaped(unsigned int m, unsigned int n, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0,
+                                                                        bool lhs_interleave, bool rhs_interleave)
+{
+    GEMMLHSMatrixInfo lhs_info;
+    GEMMRHSMatrixInfo rhs_info;
+
+    // Configure GEMMLHSMatrixInfo
+    lhs_info.m0         = m0;
+    lhs_info.k0         = k0;
+    lhs_info.v0         = ((m / (lhs_info.m0 * v0)) == 0) ? 1 : v0;
+    lhs_info.interleave = lhs_interleave;
+    lhs_info.transpose  = false;
+
+    // Configure GEMMRHSMatrixInfo
+    rhs_info.n0         = n0;
+    rhs_info.k0         = lhs_info.k0;
+    rhs_info.h0         = ((n / (rhs_info.n0 * h0)) == 0) ? 1 : h0;
+    rhs_info.interleave = rhs_interleave;
+    rhs_info.transpose  = true;
+
+    return std::make_pair(lhs_info, rhs_info);
+}
+
+} // namespace
+
 std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedConfigurationBifrost::configure(unsigned int m, unsigned int n, unsigned int k, unsigned int b, DataType data_type)
 {
-    ARM_COMPUTE_ERROR_ON(data_type != DataType::F32);
+    ARM_COMPUTE_ERROR_ON(data_type != DataType::F32 && data_type != DataType::QASYMM8);
     ARM_COMPUTE_UNUSED(data_type);
 
     const GPUTarget gpu_target = CLScheduler::get().target();
+
+    using ConfigurationFunctionExecutorPtr = std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> (CLGEMMReshapedConfigurationBifrost::*)(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+
+    // Configurations for Mali-G76
+    static std::map<DataType, ConfigurationFunctionExecutorPtr> gemm_reshaped_configs_G76 =
+    {
+        { DataType::F32, &CLGEMMReshapedConfigurationBifrost::configure_G76_f32 },
+        { DataType::QASYMM8, &CLGEMMReshapedConfigurationBifrost::configure_G76_u8 }
+    };
+
+    // Configurations for Mali-G7x
+    static std::map<DataType, ConfigurationFunctionExecutorPtr> gemm_reshaped_configs_G7x =
+    {
+        { DataType::F32, &CLGEMMReshapedConfigurationBifrost::configure_G7x_f32 },
+        { DataType::QASYMM8, &CLGEMMReshapedConfigurationBifrost::configure_G7x_u8 }
+    };
+
     switch(gpu_target)
     {
         case GPUTarget::G76:
-            return configure_G76_f32(m, n, k, b);
+            return (this->*gemm_reshaped_configs_G76[data_type])(m, n, k, b);
         default:
-            return configure_G7x_f32(m, n, k, b);
+            return (this->*gemm_reshaped_configs_G7x[data_type])(m, n, k, b);
     }
 }
 
@@ -52,43 +96,43 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedConfigurationBifro
     ARM_COMPUTE_UNUSED(k);
     ARM_COMPUTE_UNUSED(b);
 
-    GEMMLHSMatrixInfo lhs_info;
-    GEMMRHSMatrixInfo rhs_info;
-
     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;
+        return configure_gemm_reshaped(m, n, 4, 2, 8, 16, 16, true, false);
     }
     else
     {
-        // Configure GEMMLHSMatrixInfo
-        lhs_info.m0         = 5;
-        lhs_info.k0         = 4;
-        lhs_info.v0         = lhs_info.m0 * 2 < m ? 1 : 2;
-        lhs_info.interleave = false;
-        lhs_info.transpose  = false;
-
-        // Configure GEMMRHSMatrixInfo
-        rhs_info.n0         = 4;
-        rhs_info.k0         = lhs_info.k0;
-        rhs_info.h0         = rhs_info.n0 * 16 < n ? 2 : 16;
-        rhs_info.interleave = true;
-        rhs_info.transpose  = true;
+        return configure_gemm_reshaped(m, n, 5, 4, 4, 2, 16, false, true);
     }
+}
 
-    return std::make_pair(lhs_info, rhs_info);
+std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedConfigurationBifrost::configure_G7x_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
+{
+    ARM_COMPUTE_UNUSED(k);
+    ARM_COMPUTE_UNUSED(b);
+
+    if(dot8_supported(CLKernelLibrary::get().get_device()))
+    {
+        if(n <= 4)
+        {
+            return configure_gemm_reshaped(m, n, 4, 2, 16, 2, 2, true, false);
+        }
+        else
+        {
+            return configure_gemm_reshaped(m, n, 4, 4, 16, 2, 2, true, false);
+        }
+    }
+    else
+    {
+        if(n <= 4)
+        {
+            return configure_gemm_reshaped(m, n, 4, 2, 8, 2, 2, true, false);
+        }
+        else
+        {
+            return configure_gemm_reshaped(m, n, 6, 4, 4, 2, 2, true, true);
+        }
+    }
 }
 
 std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedConfigurationBifrost::configure_G76_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
@@ -96,43 +140,29 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedConfigurationBifro
     ARM_COMPUTE_UNUSED(k);
     ARM_COMPUTE_UNUSED(b);
 
-    GEMMLHSMatrixInfo lhs_info;
-    GEMMRHSMatrixInfo rhs_info;
-
     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;
+        return configure_gemm_reshaped(m, n, 4, 2, 8, 16, 16, true, false);
     }
     else
     {
-        // Configure GEMMLHSMatrixInfo
-        lhs_info.m0         = 4;
-        lhs_info.k0         = 2;
-        lhs_info.v0         = lhs_info.m0 * 8 < m ? 2 : 8;
-        lhs_info.interleave = false;
-        lhs_info.transpose  = false;
-
-        // Configure GEMMRHSMatrixInfo
-        rhs_info.n0         = 4;
-        rhs_info.k0         = lhs_info.k0;
-        rhs_info.h0         = rhs_info.n0 * 16 < n ? 2 : 16;
-        rhs_info.interleave = false;
-        rhs_info.transpose  = true;
+        return configure_gemm_reshaped(m, n, 4, 4, 2, 8, 16, false, false);
     }
+}
 
-    return std::make_pair(lhs_info, rhs_info);
+std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedConfigurationBifrost::configure_G76_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
+{
+    ARM_COMPUTE_UNUSED(k);
+    ARM_COMPUTE_UNUSED(b);
+
+    if(n <= 4)
+    {
+        return configure_gemm_reshaped(m, n, 4, 2, 16, 4, 1, false, false);
+    }
+    else
+    {
+        return configure_gemm_reshaped(m, n, 4, 4, 16, 2, 2, false, true);
+    }
 }
 } // namespace cl_gemm
 } // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/CL/GEMMLowpMatrixMultiplyReshaped.cpp b/tests/validation/CL/GEMMLowpMatrixMultiplyReshaped.cpp
new file mode 100644
index 0000000000..62b0d02373
--- /dev/null
+++ b/tests/validation/CL/GEMMLowpMatrixMultiplyReshaped.cpp
@@ -0,0 +1,199 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.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 "tests/CL/CLAccessor.h"
+#include "tests/CL/Helper.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/GEMMLowpFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+using namespace arm_compute::misc::shape_calculator;
+
+// Create function for CLGEMMReshapeLHSMatrixKernel
+using CLGEMMReshapeLHSMatrix = CLSynthetizeFunction<CLGEMMReshapeLHSMatrixKernel>;
+
+// Create function for CLGEMMReshapeRHSMatrixKernel
+using CLGEMMReshapeRHSMatrix = CLSynthetizeFunction<CLGEMMReshapeRHSMatrixKernel>;
+
+// Create function for CLGEMMMatrixMultiplyReshapedKernel
+using CLGEMMLowpMatrixMultiplyReshaped = CLSynthetizeFunction<CLGEMMLowpMatrixMultiplyReshapedKernel>;
+
+// Fixture for CLGEMMLowpMatrixMultiplyReshaped
+using CLGEMMLowpMatrixMultiplyReshapedFixture = GEMMLowpMatrixMultiplyReshapedValidationFixture<CLTensor, CLAccessor, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMLowpMatrixMultiplyReshaped>;
+
+// Fixture for CLGEMMMatrixMultiplyReshaped3D
+using CLGEMMLowpMatrixMultiplyReshaped3DFixture =
+    GEMMLowpMatrixMultiplyReshaped3DValidationFixture<CLTensor, CLAccessor, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMLowpMatrixMultiplyReshaped>;
+
+namespace
+{
+// *INDENT-OFF*
+// clang-format off
+/** M values to test */
+const auto m_values = framework::dataset::make("M", 37);
+
+/** M_W values to test */
+const auto m_w_values = framework::dataset::make("M_W", 5);
+
+/** M_H values to test */
+const auto m_h_values = framework::dataset::make("M_H", 7);
+
+/** N values to test */
+const auto n_values = framework::dataset::make("N", 51);
+
+/** K values to test */
+const auto k_values = framework::dataset::make("K", 23);
+
+/** Batch size values to test */
+const auto b_values = framework::dataset::make("batch_size", 1, 3);
+
+/** M0 values to test - Precommit */
+const auto m0_values_precommit = framework::dataset::make("M0", {4, 6});
+
+/** N0 values to test - Precommit */
+const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 });
+
+/** K0 values to test - Precommit */
+const auto k0_values_precommit = framework::dataset::make("K0", { 4 });
+
+/** 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, 7);
+
+/** N0 values to test - Nightly */
+const auto n0_values_nightly = framework::dataset::make("N0", { 2, 3, 4, 8 });
+
+/** K0 values to test - Nightly */
+const auto k0_values_nightly = framework::dataset::make("K0", { 2, 3, 4, 8 });
+
+/** 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
+
+TEST_SUITE(CL)
+TEST_SUITE(GEMMLowpMatrixMultiplyReshaped)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyReshapedFixture, 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);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyReshapedFixture, 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);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMLowpMatrixMultiplyReshaped3DFixture, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMLowpMatrixMultiplyReshaped3DFixture, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_nightly),
+                                                                   n0_values_nightly),
+                                                                   k0_values_nightly),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // GEMMLowpMatrixMulipltyReshaped
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index 96debe0eec..836f8eddfe 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -306,6 +306,233 @@ protected:
     TensorType            _target{};
     SimpleTensor<uint8_t> _reference{};
 };
+
+template <typename TensorType, typename AccessorType, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMLowpMatrixMultiplyReshapedValidationFixture : 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)
+    {
+        // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path
+        std::uniform_int_distribution<> distribution(1, 254);
+        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::QASYMM8, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, DataType::QASYMM8, 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, 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<int32_t> 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<uint8_t> lhs{ lhs_shape, DataType::QASYMM8, 1 };
+        SimpleTensor<uint8_t> rhs{ rhs_shape, DataType::QASYMM8, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+
+        return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0);
+    }
+
+    TensorType            _target{};
+    SimpleTensor<int32_t> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMLowpMatrixMultiplyReshaped3DValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int 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;
+
+        // In case of GEMM3D, m is the product between m_w and m_h
+        const unsigned int m = m_w * m_h;
+
+        // Set the tensor shapes for LHS and RHS matrices
+        const TensorShape lhs_shape(k, m, batch_size);
+        const TensorShape rhs_shape(n, k, batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, m_h);
+        _reference = compute_reference(lhs_shape, rhs_shape, m_h);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path
+        std::uniform_int_distribution<> distribution(1, 254);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, unsigned int m_h)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, DataType::QASYMM8, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, DataType::QASYMM8, 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, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h));
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        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<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, unsigned int m_h)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape.set(0, rhs_shape[0]);
+        dst_shape.set(1, lhs_shape[1] / m_h);
+        dst_shape.set(2, m_h);
+        dst_shape.set(3, lhs_shape[2]);
+
+        // Create reference
+        SimpleTensor<uint8_t> lhs{ lhs_shape, DataType::QASYMM8, 1 };
+        SimpleTensor<uint8_t> rhs{ rhs_shape, DataType::QASYMM8, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+
+        return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0);
+    }
+
+    TensorType            _target{};
+    SimpleTensor<int32_t> _reference{};
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute