diff --git a/Android.bp b/Android.bp
index ccfb2c7..adcafa6 100644
--- a/Android.bp
+++ b/Android.bp
@@ -34,7 +34,6 @@
         "src/core/CL/cl_kernels/common/floor.cl",
         "src/core/CL/cl_kernels/common/gather.cl",
         "src/core/CL/cl_kernels/common/gemm.cl",
-        "src/core/CL/cl_kernels/common/gemm_v1.cl",
         "src/core/CL/cl_kernels/common/gemmlowp.cl",
         "src/core/CL/cl_kernels/common/gemv.cl",
         "src/core/CL/cl_kernels/common/generate_proposals.cl",
@@ -529,7 +528,6 @@
         "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFloatKernel.cpp",
         "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp",
         "src/gpu/cl/kernels/ClGemmLowpReductionKernel.cpp",
-        "src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.cpp",
         "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp",
         "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.cpp",
         "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp",
diff --git a/SConscript b/SConscript
index bcb93fd..6c58391 100644
--- a/SConscript
+++ b/SConscript
@@ -308,7 +308,6 @@
                        'src/core/CL/cl_kernels/common/gather.cl',
                        'src/core/CL/cl_kernels/common/gemm.cl',
                        'src/core/CL/cl_kernels/common/gemv.cl',
-                       'src/core/CL/cl_kernels/common/gemm_v1.cl',
                        'src/core/CL/cl_kernels/common/gemmlowp.cl',
                        'src/core/CL/cl_kernels/common/generate_proposals.cl',
                        'src/core/CL/cl_kernels/common/generate_proposals_quantized.cl',
diff --git a/arm_compute/runtime/CL/CLTypes.h b/arm_compute/runtime/CL/CLTypes.h
index cf0486c..bba25c6 100644
--- a/arm_compute/runtime/CL/CLTypes.h
+++ b/arm_compute/runtime/CL/CLTypes.h
@@ -30,18 +30,8 @@
 /** OpenCL GEMM kernel types */
 enum class CLGEMMKernelType
 {
-    /** Native GEMM kernel with fixed block size.
-     * @note Temporary variant to keep compatibility with the old implementation.
-     * @note This variant will be deprecated in favor of a new and configurable NATIVE variant
-     */
-    NATIVE_V1,
     /** Native GEMM kernel with configurable block size.*/
     NATIVE,
-    /** Reshaped GEMM kernel where both lhs and rhs matrices are reshaped. Fixed block size fixed.
-     * @note Temporary variant to keep compatibility with the old implementation.
-     * @note This variant will be deprecated in favor of RESHAPED
-     */
-    RESHAPED_V1,
     /** Reshaped GEMM kernel where both lhs and rhs matrices are reshaped. Configurable reshape and block size */
     RESHAPED,
     /** Reshaped GEMM kernel where only the rhs matrix is reshaped. Configurable reshape and block size */
diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index 9235a85..2947b48 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -36,7 +36,7 @@
  *
  *  -# @ref opencl::kernels::ClIm2ColKernel (called when the input comes from a convolutional layer)
  *  -# @ref CLTranspose (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
- *  -# @ref opencl::kernels::ClGemmMatrixMultiplyKernel or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
+ *  -# @ref opencl::ClGemm or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
  *
  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
  */
diff --git a/filelist.json b/filelist.json
index bcc7ecb..5a577b9 100644
--- a/filelist.json
+++ b/filelist.json
@@ -476,7 +476,6 @@
           "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel.cpp",
           "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFloatKernel.cpp",
           "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp",
-          "src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.cpp",
           "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp",
           "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.cpp",
           "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp",
diff --git a/src/core/CL/cl_kernels/common/gemm.cl b/src/core/CL/cl_kernels/common/gemm.cl
index 87921f5..431c97b 100644
--- a/src/core/CL/cl_kernels/common/gemm.cl
+++ b/src/core/CL/cl_kernels/common/gemm.cl
@@ -4141,6 +4141,7 @@
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0)    c0=0,c1=0,c2=0,... c(M0-1)=0;
 
     int i = 0;
+#if K0 > 1
     for(; i <= (K - K0); i += K0)
     {
         // Supported cases (M0, K0):
@@ -4186,7 +4187,7 @@
         lhs_offset += K0 * sizeof(DATA_TYPE);
         rhs_offset += K0 * rhs_stride_y;
     }
-
+#endif // K0 > 1
     // Left-over accumulations
     for(; i < K; ++i)
     {
@@ -4292,10 +4293,6 @@
 
     // Store output block
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-
-#undef RHS_BLOCK_SIZE
-#undef RHS_OFFSET_X
-#undef RHS_STEP_X
 }
 #endif // defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE)
 
diff --git a/src/core/CL/cl_kernels/common/gemm_v1.cl b/src/core/CL/cl_kernels/common/gemm_v1.cl
deleted file mode 100644
index a136a1b..0000000
--- a/src/core/CL/cl_kernels/common/gemm_v1.cl
+++ /dev/null
@@ -1,3243 +0,0 @@
-/*
- * Copyright (c) 2020-2021 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 "gemm_helpers.h"
-#include "repeat.h"
-
-#if defined(M) && defined(N) && defined(K) && defined(H0) && defined(V0) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) && defined(IN1_DIM_X)
-/** This OpenCL kernel is optimised for Midgard. It computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1)
- *
- * @note The number of rows of destination matrix must be passed at compile time using -DM
- * @note The number of columns of the destination matrix must be passed at compile time using -DN
- * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK
- * @note The number of columns of the reshaped rhs matrix must be passed at compile time using -DIN1_DIM_X
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2)
- * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2)
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the output has to be reinterpreted as a 3D tensor (e.g. 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 matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  cross_plane_pad                    (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_interleaved_transposed_f32(IMAGE_DECLARATION(src0),
-                                                 IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                                 IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                                 IMAGE_DECLARATION(dst),
-                                                 uint src0_stride_z,
-                                                 uint src1_stride_z,
-#if defined(BETA)
-                                                 uint src2_stride_z,
-#endif //defined(BETA)
-                                                 uint dst_stride_z
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                                 ,
-                                                 uint cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                                )
-{
-    int x = get_global_id(0) / H0;
-    int y = get_global_id(1) / V0;
-    int z = get_global_id(2);
-
-    // Offset
-    const int offset_row_a = (get_global_id(1) % V0) * 4;
-    const int offset_row_b = (get_global_id(0) % H0) * 4;
-
-    // src_addr_a = address of matrix A
-    // src_addr_b = address of matrix B
-    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
-    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src1_addr_in_bytes += z * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    __global float *src_addr_a = (__global float *)(src0_ptr + src0_addr_in_bytes);
-    __global float *src_addr_b = (__global float *)(src1_ptr + src1_addr_in_bytes);
-
-    // Compute end row address for matrix B
-    __global float *src_end_addr_b = src_addr_b + IN1_DIM_X;
-
-    src_addr_a += offset_row_a;
-    src_addr_b += offset_row_b;
-
-    // Reset accumulators
-    float4 c0 = 0.0f;
-    float4 c1 = 0.0f;
-    float4 c2 = 0.0f;
-    float4 c3 = 0.0f;
-
-    for(; src_addr_b <= (src_end_addr_b - (int)(8 * H0)); src_addr_a += 8 * V0, src_addr_b += 8 * H0)
-    {
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        float4 a0 = vload4(0, src_addr_a);
-        float4 b0 = vload4(0, src_addr_b);
-
-        c0 += (float4)a0.s0 * b0;
-        c1 += (float4)a0.s1 * b0;
-        c2 += (float4)a0.s2 * b0;
-        c3 += (float4)a0.s3 * b0;
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = vload4(0, src_addr_a + 4 * V0);
-        b0 = vload4(0, src_addr_b + 4 * H0);
-
-        c0 += (float4)a0.s0 * b0;
-        c1 += (float4)a0.s1 * b0;
-        c2 += (float4)a0.s2 * b0;
-        c3 += (float4)a0.s3 * b0;
-    }
-
-    for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * V0, src_addr_b += 4 * H0)
-    {
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        float4 a0 = vload4(0, src_addr_a);
-        float4 b0 = vload4(0, src_addr_b);
-
-        c0 += (float4)a0.s0 * b0;
-        c1 += (float4)a0.s1 * b0;
-        c2 += (float4)a0.s2 * b0;
-        c3 += (float4)a0.s3 * b0;
-    }
-
-    // Compute destination address
-    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
-
-    // Compute dst address
-    __global uchar *dst_addr = offset(&dst, 0, 0);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(4, float, c, ALPHA);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float));
-
-    LOAD_BLOCK(1, 4, float, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, float, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(4, c, bias0);
-
-#else // defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id(
-                                    2) * src2_stride_z;
-
-    LOAD_BLOCK(4, 4, float, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(4, float, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    ADD_BLOCK(4, c, bias);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(4, ACTIVATION_TYPE, float, VEC_SIZE, c, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store 4x4 block
-    const bool cond_y = ((get_global_id(1) + 1) * 4 >= M);
-    const bool cond_x = ((get_global_id(0) + 1) * 4 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(4, 4, float, c, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-
-/** This OpenCL kernel is optimized for Bifrost and tt computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1)
- *
- * @note The number of rows of destination matrix must be passed at compile time using -DM
- * @note The number of columns of the destination matrix must be passed at compile time using -DN
- * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2)
- * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2)
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the output has to be reinterpreted as a 3D tensor (e.g. 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 matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  cross_plane_pad                    (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_interleaved_transposed_f32_bifrost(IMAGE_DECLARATION(src0),
-                                                         IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                                         IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                                         IMAGE_DECLARATION(dst),
-                                                         uint src0_stride_z,
-                                                         uint src1_stride_z,
-#if defined(BETA)
-                                                         uint src2_stride_z,
-#endif //defined(BETA)
-                                                         uint dst_stride_z
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                                         ,
-                                                         uint cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                                        )
-{
-    int x = get_global_id(0) / H0;
-    int y = get_global_id(1) / V0;
-    int z = get_global_id(2);
-
-    // Offset
-    const int offset_row_a = (get_global_id(1) % V0) * 4;
-    const int offset_row_b = (get_global_id(0) % H0) * 4;
-
-    // src_addr_a = address of matrix A
-    // src_addr_b = address of matrix B
-    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
-    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src1_addr_in_bytes += z * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    __global float *src_addr_a = (__global float *)(src0_ptr + src0_addr_in_bytes);
-    __global float *src_addr_b = (__global float *)(src1_ptr + src1_addr_in_bytes);
-
-    src_addr_a += offset_row_a;
-    src_addr_b += offset_row_b;
-
-    // Reset accumulators
-    float4 c0 = 0.0f;
-    float4 c1 = 0.0f;
-    float4 c2 = 0.0f;
-    float4 c3 = 0.0f;
-
-    int i = 0;
-    for(; i <= (int)(K - 4); i += 4)
-    {
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        float4 a0 = vload4(0, src_addr_a);
-        float4 b0 = vload4(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 4 * H0;
-
-        c0.s0 = fma(a0.s0, b0.s0, c0.s0);
-        c0.s1 = fma(a0.s0, b0.s1, c0.s1);
-        c0.s2 = fma(a0.s0, b0.s2, c0.s2);
-        c0.s3 = fma(a0.s0, b0.s3, c0.s3);
-
-        c1.s0 = fma(a0.s1, b0.s0, c1.s0);
-        c1.s1 = fma(a0.s1, b0.s1, c1.s1);
-        c1.s2 = fma(a0.s1, b0.s2, c1.s2);
-        c1.s3 = fma(a0.s1, b0.s3, c1.s3);
-
-        c2.s0 = fma(a0.s2, b0.s0, c2.s0);
-        c2.s1 = fma(a0.s2, b0.s1, c2.s1);
-        c2.s2 = fma(a0.s2, b0.s2, c2.s2);
-        c2.s3 = fma(a0.s2, b0.s3, c2.s3);
-
-        c3.s0 = fma(a0.s3, b0.s0, c3.s0);
-        c3.s1 = fma(a0.s3, b0.s1, c3.s1);
-        c3.s2 = fma(a0.s3, b0.s2, c3.s2);
-        c3.s3 = fma(a0.s3, b0.s3, c3.s3);
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = vload4(0, src_addr_a);
-        b0 = vload4(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 4 * H0;
-
-        c0.s0 = fma(a0.s0, b0.s0, c0.s0);
-        c0.s1 = fma(a0.s0, b0.s1, c0.s1);
-        c0.s2 = fma(a0.s0, b0.s2, c0.s2);
-        c0.s3 = fma(a0.s0, b0.s3, c0.s3);
-
-        c1.s0 = fma(a0.s1, b0.s0, c1.s0);
-        c1.s1 = fma(a0.s1, b0.s1, c1.s1);
-        c1.s2 = fma(a0.s1, b0.s2, c1.s2);
-        c1.s3 = fma(a0.s1, b0.s3, c1.s3);
-
-        c2.s0 = fma(a0.s2, b0.s0, c2.s0);
-        c2.s1 = fma(a0.s2, b0.s1, c2.s1);
-        c2.s2 = fma(a0.s2, b0.s2, c2.s2);
-        c2.s3 = fma(a0.s2, b0.s3, c2.s3);
-
-        c3.s0 = fma(a0.s3, b0.s0, c3.s0);
-        c3.s1 = fma(a0.s3, b0.s1, c3.s1);
-        c3.s2 = fma(a0.s3, b0.s2, c3.s2);
-        c3.s3 = fma(a0.s3, b0.s3, c3.s3);
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = vload4(0, src_addr_a);
-        b0 = vload4(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 4 * H0;
-
-        c0.s0 = fma(a0.s0, b0.s0, c0.s0);
-        c0.s1 = fma(a0.s0, b0.s1, c0.s1);
-        c0.s2 = fma(a0.s0, b0.s2, c0.s2);
-        c0.s3 = fma(a0.s0, b0.s3, c0.s3);
-
-        c1.s0 = fma(a0.s1, b0.s0, c1.s0);
-        c1.s1 = fma(a0.s1, b0.s1, c1.s1);
-        c1.s2 = fma(a0.s1, b0.s2, c1.s2);
-        c1.s3 = fma(a0.s1, b0.s3, c1.s3);
-
-        c2.s0 = fma(a0.s2, b0.s0, c2.s0);
-        c2.s1 = fma(a0.s2, b0.s1, c2.s1);
-        c2.s2 = fma(a0.s2, b0.s2, c2.s2);
-        c2.s3 = fma(a0.s2, b0.s3, c2.s3);
-
-        c3.s0 = fma(a0.s3, b0.s0, c3.s0);
-        c3.s1 = fma(a0.s3, b0.s1, c3.s1);
-        c3.s2 = fma(a0.s3, b0.s2, c3.s2);
-        c3.s3 = fma(a0.s3, b0.s3, c3.s3);
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = vload4(0, src_addr_a);
-        b0 = vload4(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 4 * H0;
-
-        c0.s0 = fma(a0.s0, b0.s0, c0.s0);
-        c0.s1 = fma(a0.s0, b0.s1, c0.s1);
-        c0.s2 = fma(a0.s0, b0.s2, c0.s2);
-        c0.s3 = fma(a0.s0, b0.s3, c0.s3);
-
-        c1.s0 = fma(a0.s1, b0.s0, c1.s0);
-        c1.s1 = fma(a0.s1, b0.s1, c1.s1);
-        c1.s2 = fma(a0.s1, b0.s2, c1.s2);
-        c1.s3 = fma(a0.s1, b0.s3, c1.s3);
-
-        c2.s0 = fma(a0.s2, b0.s0, c2.s0);
-        c2.s1 = fma(a0.s2, b0.s1, c2.s1);
-        c2.s2 = fma(a0.s2, b0.s2, c2.s2);
-        c2.s3 = fma(a0.s2, b0.s3, c2.s3);
-
-        c3.s0 = fma(a0.s3, b0.s0, c3.s0);
-        c3.s1 = fma(a0.s3, b0.s1, c3.s1);
-        c3.s2 = fma(a0.s3, b0.s2, c3.s2);
-        c3.s3 = fma(a0.s3, b0.s3, c3.s3);
-    }
-
-    for(; i < (int)K; ++i)
-    {
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        float4 a0 = vload4(0, src_addr_a);
-        float4 b0 = vload4(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 4 * H0;
-
-        c0.s0 = fma(a0.s0, b0.s0, c0.s0);
-        c0.s1 = fma(a0.s0, b0.s1, c0.s1);
-        c0.s2 = fma(a0.s0, b0.s2, c0.s2);
-        c0.s3 = fma(a0.s0, b0.s3, c0.s3);
-
-        c1.s0 = fma(a0.s1, b0.s0, c1.s0);
-        c1.s1 = fma(a0.s1, b0.s1, c1.s1);
-        c1.s2 = fma(a0.s1, b0.s2, c1.s2);
-        c1.s3 = fma(a0.s1, b0.s3, c1.s3);
-
-        c2.s0 = fma(a0.s2, b0.s0, c2.s0);
-        c2.s1 = fma(a0.s2, b0.s1, c2.s1);
-        c2.s2 = fma(a0.s2, b0.s2, c2.s2);
-        c2.s3 = fma(a0.s2, b0.s3, c2.s3);
-
-        c3.s0 = fma(a0.s3, b0.s0, c3.s0);
-        c3.s1 = fma(a0.s3, b0.s1, c3.s1);
-        c3.s2 = fma(a0.s3, b0.s2, c3.s2);
-        c3.s3 = fma(a0.s3, b0.s3, c3.s3);
-    }
-
-    // Compute destination address
-    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
-
-    // Compute dst address
-    __global uchar *dst_addr = offset(&dst, 0, 0);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(4, float, c, ALPHA);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float));
-
-    LOAD_BLOCK(1, 4, float, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, float, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(4, c, bias0);
-
-#else // defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id(
-                                    2) * src2_stride_z;
-
-    LOAD_BLOCK(4, 4, float, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(4, float, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    ADD_BLOCK(4, c, bias);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(4, ACTIVATION_TYPE, float, VEC_SIZE, c, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store 4x4 block
-    const bool cond_y = ((get_global_id(1) + 1) * 4 >= M);
-    const bool cond_x = ((get_global_id(0) + 1) * 4 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(4, 4, float, c, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-
-#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
-/** This OpenCL kernel computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1)
- *
- * @note The number of rows of destination matrix must be passed at compile time using -DM
- * @note The number of columns of the destination matrix must be passed at compile time using -DN
- * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK
- * @note The number of columns of the reshaped rhs matrix must be passed at compile time using -DIN1_DIM_X
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2)
- * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2)
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the output has to be reinterpreted as a 3D tensor (e.g. 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 matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  cross_plane_pad                    (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_interleaved_transposed_f16(IMAGE_DECLARATION(src0),
-                                                 IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                                 IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                                 IMAGE_DECLARATION(dst),
-                                                 uint src0_stride_z,
-                                                 uint src1_stride_z,
-#if defined(BETA)
-                                                 uint src2_stride_z,
-#endif //defined(BETA)
-                                                 uint dst_stride_z
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                                 ,
-                                                 uint cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                                )
-{
-    int x = get_global_id(0) / H0;
-    int y = get_global_id(1) / V0;
-    int z = get_global_id(2);
-
-    // Offset
-    const int offset_row_a = (get_global_id(1) % V0) * 4;
-    const int offset_row_b = (get_global_id(0) % H0) * 8;
-
-    // src_addr_a = address of matrix A
-    // src_addr_b = address of matrix B
-    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
-    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src1_addr_in_bytes += z * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes);
-    __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes);
-
-    // Compute end row address for matrix B
-    __global half *src_end_addr_b = src_addr_b + IN1_DIM_X;
-
-    src_addr_a += offset_row_a;
-    src_addr_b += offset_row_b;
-
-    // Reset accumulators
-    half8 c0 = 0.0f;
-    half8 c1 = 0.0f;
-    half8 c2 = 0.0f;
-    half8 c3 = 0.0f;
-
-    for(; src_addr_b <= (src_end_addr_b - (int)(16 * H0)); src_addr_a += 8 * V0, src_addr_b += 16 * H0)
-    {
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        half4 a0 = vload4(0, src_addr_a);
-        half8 b0 = vload8(0, src_addr_b);
-
-        c0 += (half8)a0.s0 * b0;
-        c1 += (half8)a0.s1 * b0;
-        c2 += (half8)a0.s2 * b0;
-        c3 += (half8)a0.s3 * b0;
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = vload4(0, src_addr_a + 4 * V0);
-        b0 = vload8(0, src_addr_b + 8 * H0);
-
-        c0 += (half8)a0.s0 * b0;
-        c1 += (half8)a0.s1 * b0;
-        c2 += (half8)a0.s2 * b0;
-        c3 += (half8)a0.s3 * b0;
-    }
-
-    for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * V0, src_addr_b += 8 * H0)
-    {
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        half4 a0 = vload4(0, src_addr_a);
-        half8 b0 = vload8(0, src_addr_b);
-
-        c0 += (half8)a0.s0 * b0;
-        c1 += (half8)a0.s1 * b0;
-        c2 += (half8)a0.s2 * b0;
-        c3 += (half8)a0.s3 * b0;
-    }
-
-    // Compute destination address
-    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
-
-    // Compute dst address
-    __global uchar *dst_addr = offset(&dst, 0, 0);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(4, half, c, ALPHA);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half));
-
-    LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, half, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(4, c, bias0);
-
-#else // defined(BROADCAST_BIAS)
-
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id(
-                                    2) * src2_stride_z;
-
-    LOAD_BLOCK(4, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(4, half, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    ADD_BLOCK(4, c, bias);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(4, ACTIVATION_TYPE, half, VEC_SIZE, c, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store 4x8 block
-    const bool cond_y = ((get_global_id(1) + 1) * 4 >= M);
-    const bool cond_x = ((get_global_id(0) + 1) * 8 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(4, 8, half, c, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-
-/** This OpenCL kernel computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) while accumulating the result in a 32 floating point variable.
- *
- * @note The number of rows of destination matrix must be passed at compile time using -DM
- * @note The number of columns of the destination matrix must be passed at compile time using -DN
- * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK
- * @note The number of columns of the reshaped rhs matrix must be passed at compile time using -DIN1_DIM_X
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2)
- * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2)
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the output has to be reinterpreted as a 3D tensor (e.g. 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 matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  cross_plane_pad                    (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_interleaved_transposed_f16_acc32(IMAGE_DECLARATION(src0),
-                                                       IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                                       IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                                       IMAGE_DECLARATION(dst),
-                                                       uint src0_stride_z,
-                                                       uint src1_stride_z,
-#if defined(BETA)
-                                                       uint src2_stride_z,
-#endif //defined(BETA)
-                                                       uint dst_stride_z
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                                       ,
-                                                       uint cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                                      )
-{
-    int x = get_global_id(0) / H0;
-    int y = get_global_id(1) / V0;
-    int z = get_global_id(2);
-
-    // Offset
-    const int offset_row_a = (get_global_id(1) % V0) * 4;
-    const int offset_row_b = (get_global_id(0) % H0) * 8;
-
-    // src_addr_a = address of matrix A
-    // src_addr_b = address of matrix B
-    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
-    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src1_addr_in_bytes += z * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes);
-    __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes);
-
-    // Compute end row address for matrix B
-    __global half *src_end_addr_b = src_addr_b + IN1_DIM_X;
-
-    src_addr_a += offset_row_a;
-    src_addr_b += offset_row_b;
-
-    // Reset accumulators
-    float8 c0 = 0.0f;
-    float8 c1 = 0.0f;
-    float8 c2 = 0.0f;
-    float8 c3 = 0.0f;
-
-    for(; src_addr_b <= (src_end_addr_b - (int)(16 * H0)); src_addr_a += 8 * V0, src_addr_b += 16 * H0)
-    {
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        float4 a0 = convert_float4(vload4(0, src_addr_a));
-        float8 b0 = convert_float8(vload8(0, src_addr_b));
-
-        c0 += (float8)a0.s0 * b0;
-        c1 += (float8)a0.s1 * b0;
-        c2 += (float8)a0.s2 * b0;
-        c3 += (float8)a0.s3 * b0;
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = convert_float4(vload4(0, src_addr_a + 4 * V0));
-        b0 = convert_float8(vload8(0, src_addr_b + 8 * H0));
-
-        c0 += (float8)a0.s0 * b0;
-        c1 += (float8)a0.s1 * b0;
-        c2 += (float8)a0.s2 * b0;
-        c3 += (float8)a0.s3 * b0;
-    }
-
-    for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * V0, src_addr_b += 8 * H0)
-    {
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        float4 a0 = convert_float4(vload4(0, src_addr_a));
-        float8 b0 = convert_float8(vload8(0, src_addr_b));
-
-        c0 += (float8)a0.s0 * b0;
-        c1 += (float8)a0.s1 * b0;
-        c2 += (float8)a0.s2 * b0;
-        c3 += (float8)a0.s3 * b0;
-    }
-
-    // Compute destination address
-    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
-
-    // Compute dst address
-    __global uchar *dst_addr = offset(&dst, 0, 0);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(4, float, c, ALPHA);
-#endif // defined(ALPHA)
-
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half));
-
-    LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-    float8 bias_f0 = convert_float8(bias0);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, float, bias_f, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(4, c, bias_f0);
-
-#else // defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id(
-                                    2) * src2_stride_z;
-
-    LOAD_BLOCK(4, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-    float8 bias_f0 = convert_float8(bias0);
-    float8 bias_f1 = convert_float8(bias1);
-    float8 bias_f2 = convert_float8(bias2);
-    float8 bias_f3 = convert_float8(bias3);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(4, float, bias_f, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    ADD_BLOCK(4, c, bias_f);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-    half8 c_h0 = convert_half8(c0);
-    half8 c_h1 = convert_half8(c1);
-    half8 c_h2 = convert_half8(c2);
-    half8 c_h3 = convert_half8(c3);
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(4, ACTIVATION_TYPE, half, VEC_SIZE, c_h, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store 4x8 block
-    const bool cond_y = ((get_global_id(1) + 1) * 4 >= M);
-    const bool cond_x = ((get_global_id(0) + 1) * 8 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(4, 8, half, c_h, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-
-/** This OpenCL kernel optimized for Bifrost architectures computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1)
- *
- * @note The number of rows of destination matrix must be passed at compile time using -DM
- * @note The number of columns of the destination matrix must be passed at compile time using -DN
- * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2)
- * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2)
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the output has to be reinterpreted as a 3D tensor (e.g. 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 matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  cross_plane_pad                    (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_interleaved_transposed_f16_bifrost(IMAGE_DECLARATION(src0),
-                                                         IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                                         IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                                         IMAGE_DECLARATION(dst),
-                                                         uint src0_stride_z,
-                                                         uint src1_stride_z,
-#if defined(BETA)
-                                                         uint src2_stride_z,
-#endif //defined(BETA)
-                                                         uint dst_stride_z
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                                         ,
-                                                         uint cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                                        )
-{
-    int x = get_global_id(0) / H0;
-    int y = get_global_id(1) / V0;
-    int z = get_global_id(2);
-
-    // Offset
-    const int offset_row_a = (get_global_id(1) % V0) * 4;
-    const int offset_row_b = (get_global_id(0) % H0) * 8;
-
-    // src_addr_a = address of matrix A
-    // src_addr_b = address of matrix B
-    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
-    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src1_addr_in_bytes += z * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes);
-    __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes);
-
-    src_addr_a += offset_row_a;
-    src_addr_b += offset_row_b;
-
-    // Reset accumulators
-    half8 c0 = 0.0f;
-    half8 c1 = 0.0f;
-    half8 c2 = 0.0f;
-    half8 c3 = 0.0f;
-
-    int i = 0;
-    for(; i <= (int)(K - 4); i += 4)
-    {
-#if V0 == 1
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        half8 a0 = vload8(0, src_addr_a);
-        half8 b0 = vload8(0, src_addr_b);
-
-        src_addr_a += 8 * V0;
-        src_addr_b += 8 * H0;
-
-        c0 = fma((half8)a0.s0, b0, c0);
-        c1 = fma((half8)a0.s1, b0, c1);
-        c2 = fma((half8)a0.s2, b0, c2);
-        c3 = fma((half8)a0.s3, b0, c3);
-
-        // Load values from matrix B (transposed)
-        b0 = vload8(0, src_addr_b);
-
-        src_addr_b += 8 * H0;
-
-        c0 = fma((half8)a0.s4, b0, c0);
-        c1 = fma((half8)a0.s5, b0, c1);
-        c2 = fma((half8)a0.s6, b0, c2);
-        c3 = fma((half8)a0.s7, b0, c3);
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = vload8(0, src_addr_a);
-        b0 = vload8(0, src_addr_b);
-
-        src_addr_a += 8 * V0;
-        src_addr_b += 8 * H0;
-
-        c0 = fma((half8)a0.s0, b0, c0);
-        c1 = fma((half8)a0.s1, b0, c1);
-        c2 = fma((half8)a0.s2, b0, c2);
-        c3 = fma((half8)a0.s3, b0, c3);
-
-        // Load values from matrix B (transposed)
-        b0 = vload8(0, src_addr_b);
-
-        src_addr_b += 8 * H0;
-
-        c0 = fma((half8)a0.s4, b0, c0);
-        c1 = fma((half8)a0.s5, b0, c1);
-        c2 = fma((half8)a0.s6, b0, c2);
-        c3 = fma((half8)a0.s7, b0, c3);
-#else  // V0 == 1
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        half4 a0 = vload4(0, src_addr_a);
-        half8 b0 = vload8(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 8 * H0;
-
-        c0 = fma((half8)a0.s0, b0, c0);
-        c1 = fma((half8)a0.s1, b0, c1);
-        c2 = fma((half8)a0.s2, b0, c2);
-        c3 = fma((half8)a0.s3, b0, c3);
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = vload4(0, src_addr_a);
-        b0 = vload8(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 8 * H0;
-
-        c0 = fma((half8)a0.s0, b0, c0);
-        c1 = fma((half8)a0.s1, b0, c1);
-        c2 = fma((half8)a0.s2, b0, c2);
-        c3 = fma((half8)a0.s3, b0, c3);
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = vload4(0, src_addr_a);
-        b0 = vload8(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 8 * H0;
-
-        c0 = fma((half8)a0.s0, b0, c0);
-        c1 = fma((half8)a0.s1, b0, c1);
-        c2 = fma((half8)a0.s2, b0, c2);
-        c3 = fma((half8)a0.s3, b0, c3);
-
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        a0 = vload4(0, src_addr_a);
-        b0 = vload8(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 8 * H0;
-
-        c0 = fma((half8)a0.s0, b0, c0);
-        c1 = fma((half8)a0.s1, b0, c1);
-        c2 = fma((half8)a0.s2, b0, c2);
-        c3 = fma((half8)a0.s3, b0, c3);
-#endif // V0 == 1
-    }
-
-    for(; i < (int)K; ++i)
-    {
-        // Load values from matrix A (interleaved) and matrix B (transposed)
-        half4 a0 = vload4(0, src_addr_a);
-        half8 b0 = vload8(0, src_addr_b);
-
-        src_addr_a += 4 * V0;
-        src_addr_b += 8 * H0;
-
-        c0 = fma((half8)a0.s0, b0, c0);
-        c1 = fma((half8)a0.s1, b0, c1);
-        c2 = fma((half8)a0.s2, b0, c2);
-        c3 = fma((half8)a0.s3, b0, c3);
-    }
-
-    // Compute destination address
-    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
-
-    // Compute dst address
-    __global uchar *dst_addr = offset(&dst, 0, 0);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(4, half, c, ALPHA);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half));
-
-    LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, half, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(4, c, bias0);
-
-#else // defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id(
-                                    2) * src2_stride_z;
-
-    LOAD_BLOCK(4, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(4, half, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    ADD_BLOCK(4, c, bias);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(4, ACTIVATION_TYPE, half, VEC_SIZE, c, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store 4x8 block
-    const bool cond_y = ((get_global_id(1) + 1) * 4 >= M);
-    const bool cond_x = ((get_global_id(0) + 1) * 8 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(4, 8, half, c, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-
-#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
-
-#endif // defined(M) && defined(N) && defined(K) && defined(H0) && defined(V0) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) && defined(IN1_DIM_X)
-
-#if defined(N) && defined(K) && defined(M0) && defined(N0) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
-#if defined(DATA_TYPE)
-#define VECTOR_TYPE VEC_DATA_TYPE(DATA_TYPE, N0)
-/** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not been reshaped.
- *
- * @note This OpenCL kernel works with floating point data types (F16/F32)
- * @note The floating point data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
- * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0
- * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
- *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
- *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
- *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
- *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
- *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16/F32
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
- * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
- * @param[in]  dst_step_x                         dst_gx_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_gx_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  src_cross_plane_pad                (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D)
- * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings in unit of elements for the output tensor (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_floating_point(IMAGE_DECLARATION(src0),
-                                     IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                     IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                     IMAGE_DECLARATION(dst),
-                                     uint src0_stride_z,
-                                     uint src1_stride_z,
-#if defined(BETA)
-                                     uint src2_stride_z,
-#endif //defined(BETA)
-                                     uint dst_stride_z
-#if defined(REINTERPRET_INPUT_AS_3D)
-                                     ,
-                                     uint src_cross_plane_pad
-#endif // REINTERPRET_INPUT_AS_3D
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                     ,
-                                     uint dst_cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                    )
-{
-    int idx = get_global_id(0) * N0;
-
-    // Compute starting address for matrix A and Matrix B
-    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
-
-    // Update address for the matrix A
-    src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y;
-
-    // Update address for the matrix B
-    src_addr.s1 += idx * sizeof(DATA_TYPE);
-
-#if defined(REINTERPRET_INPUT_AS_3D)
-    // Since we load a 2D input tile from 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 row by HEIGHT_GEMM3D
-    uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zin       = min(DEPTH_GEMM3D - 1, zin);
-
-    // Add offset due to the cross plane paddings
-    zin *= (src_cross_plane_pad * src0_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply src0_stride_z by DEPTH_GEMM3D
-    src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;
-
-#else // defined(REINTERPRET_INPUT_AS_3D)
-
-    // Add offset for batched GEMM
-    src_addr.s0 += get_global_id(2) * src0_stride_z;
-
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src_addr.s1 += get_global_id(2) * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    int end_row_vec_a = src_addr.s0 + (K * sizeof(DATA_TYPE));
-
-    VECTOR_TYPE acc0 = 0.0f;
-#if M0 > 1
-    VECTOR_TYPE acc1 = 0.0f;
-#endif // M0 > 1
-#if M0 > 2
-    VECTOR_TYPE acc2 = 0.0f;
-#endif // M0 > 2
-#if M0 > 3
-    VECTOR_TYPE acc3 = 0.0f;
-#endif // M0 > 3
-
-    for(; src_addr.s0 <= (end_row_vec_a - 2 * (int)sizeof(DATA_TYPE)); src_addr += (int2)(2 * sizeof(DATA_TYPE), 2 * src1_stride_y))
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        LOAD_BLOCK(M0, 2, DATA_TYPE, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s);
-#else // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        VEC_DATA_TYPE(DATA_TYPE, 2)
-        a0 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
-#if M0 > 1
-        VEC_DATA_TYPE(DATA_TYPE, 2)
-        a1 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // M0 > 1
-#if M0 > 2
-        VEC_DATA_TYPE(DATA_TYPE, 2)
-        a2 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // M0 > 2
-#if M0 > 3
-        VEC_DATA_TYPE(DATA_TYPE, 2)
-        a3 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // M0 > 3
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        // Load values from matrix B
-        VECTOR_TYPE b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1));
-        VECTOR_TYPE b1 = VLOAD(N0)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1 + src1_stride_y));
-
-        // Accumulate
-        acc0 += b0 * (VECTOR_TYPE)a0.s0;
-        acc0 += b1 * (VECTOR_TYPE)a0.s1;
-#if M0 > 1
-        acc1 += b0 * (VECTOR_TYPE)a1.s0;
-        acc1 += b1 * (VECTOR_TYPE)a1.s1;
-#endif // M0 > 1
-#if M0 > 2
-        acc2 += b0 * (VECTOR_TYPE)a2.s0;
-        acc2 += b1 * (VECTOR_TYPE)a2.s1;
-#endif // M0 > 2
-#if M0 > 3
-        acc3 += b0 * (VECTOR_TYPE)a3.s0;
-        acc3 += b1 * (VECTOR_TYPE)a3.s1;
-#endif // M0 > 3
-    }
-
-    for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(DATA_TYPE), src1_stride_y))
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        DATA_TYPE a0 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0));
-#if M0 > 1
-        DATA_TYPE a1 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1));
-#endif // M0 > 1
-#if M0 > 2
-        DATA_TYPE a2 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2));
-#endif // M0 > 2
-#if M0 > 3
-        DATA_TYPE a3 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3));
-#endif // M0 > 3
-#else  // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        DATA_TYPE a0 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
-#if M0 > 1
-        DATA_TYPE a1 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // M0 > 1
-#if M0 > 2
-        DATA_TYPE a2 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // M0 > 2
-#if M0 > 3
-        DATA_TYPE a3 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // M0 > 3
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        // Load values from matrix B
-        VECTOR_TYPE b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1));
-
-        // Accumulate
-        acc0 += b0 * (VECTOR_TYPE)a0;
-#if M0 > 1
-        acc1 += b0 * (VECTOR_TYPE)a1;
-#endif // M0 > 1
-#if M0 > 2
-        acc2 += b0 * (VECTOR_TYPE)a2;
-#endif // M0 > 2
-#if M0 > 3
-        acc3 += b0 * (VECTOR_TYPE)a3;
-#endif // M0 > 3
-    }
-
-    int z = get_global_id(2);
-
-    // Compute dst address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0,
-                               PARTIAL_STORE_M0)
-                               * dst_stride_y);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (dst_cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(M0, DATA_TYPE, acc, ALPHA);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
-
-    LOAD_BLOCK(1, N0, DATA_TYPE, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(M0, acc, bias0);
-
-#else // defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0,
-                                PARTIAL_STORE_M0)
-                                * src2_stride_y)
-                                + z * src2_stride_z;
-
-    LOAD_BLOCK(M0, N0, DATA_TYPE, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    ADD_BLOCK(M0, acc, bias);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, acc, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store output block
-    const bool cond_y = get_global_id(1) == 0;
-    const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, acc, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-#endif // defined(DATA_TYPE)
-
-/** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not been reshaped
- *
- * @note This OpenCL kernel works with the 32-bit floating point data type (float) and uses the fma units.
- * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0.
- * @note This kernel processed a fixed number of elements along x: -DN0=4.
- * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
- *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
- *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
- *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
- *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
- *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
- * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
- * @param[in]  dst_step_x                         dst_gx_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_gx_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  src_cross_plane_pad                (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D)
- * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_floating_point_f32_bifrost(IMAGE_DECLARATION(src0),
-                                                 IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                                 IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                                 IMAGE_DECLARATION(dst),
-                                                 uint src0_stride_z,
-                                                 uint src1_stride_z,
-#if defined(BETA)
-                                                 uint src2_stride_z,
-#endif //defined(BETA)
-                                                 uint dst_stride_z
-#if defined(REINTERPRET_INPUT_AS_3D)
-                                                 ,
-                                                 uint src_cross_plane_pad
-#endif // REINTERPRET_INPUT_AS_3D
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                                 ,
-                                                 uint dst_cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                                )
-{
-    int idx = get_global_id(0) * N0;
-
-    // Compute starting address for matrix A and matrix B
-    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
-
-    // Update address for matrix A
-    src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y;
-
-    // Update address for matrix B
-    src_addr.s1 += idx * sizeof(float);
-
-#if defined(REINTERPRET_INPUT_AS_3D)
-    // Since we load a 2D input tile from 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 row by HEIGHT_GEMM3D
-    uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zin       = min(DEPTH_GEMM3D - 1, zin);
-
-    // Add offset due to the cross plane paddings
-    zin *= (src_cross_plane_pad * src0_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply src0_stride_z by DEPTH_GEMM3D
-    src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;
-
-#else // defined(REINTERPRET_INPUT_AS_3D)
-
-    // Add offset for batched GEMM
-    src_addr.s0 += get_global_id(2) * src0_stride_z;
-
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src_addr.s1 += get_global_id(2) * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    // Initialize accumulators
-    float4 acc0 = 0.0f;
-
-#if M0 > 1
-    float4 acc1 = 0.0f;
-#endif // M0 > 1
-
-#if M0 > 2
-    float4 acc2 = 0.0f;
-#endif // M0 > 2
-
-#if M0 > 3
-    float4 acc3 = 0.0f;
-#endif // M0 > 3
-
-    // A and B src indices get incremented at the same time.
-    int i = 0;
-    for(; i <= ((int)K - 4); i += 4)
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A and matrix B
-        LOAD_BLOCK(M0, 4, float, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s);
-#else // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A and matrix B
-        float4 a0 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
-#if M0 > 1
-        float4 a1 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // M0 > 1
-#if M0 > 2
-        float4 a2 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // M0 > 2
-#if M0 > 3
-        float4 a3 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // M0 > 3
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-
-        // Multiply and accumulate
-        acc0.s0 = fma(a0.s0, b0.s0, acc0.s0);
-        acc0.s1 = fma(a0.s0, b0.s1, acc0.s1);
-        acc0.s2 = fma(a0.s0, b0.s2, acc0.s2);
-        acc0.s3 = fma(a0.s0, b0.s3, acc0.s3);
-
-#if M0 > 1
-
-        acc1.s0 = fma(a1.s0, b0.s0, acc1.s0);
-        acc1.s1 = fma(a1.s0, b0.s1, acc1.s1);
-        acc1.s2 = fma(a1.s0, b0.s2, acc1.s2);
-        acc1.s3 = fma(a1.s0, b0.s3, acc1.s3);
-
-#endif // M0 > 1
-#if M0 > 2
-
-        acc2.s0 = fma(a2.s0, b0.s0, acc2.s0);
-        acc2.s1 = fma(a2.s0, b0.s1, acc2.s1);
-        acc2.s2 = fma(a2.s0, b0.s2, acc2.s2);
-        acc2.s3 = fma(a2.s0, b0.s3, acc2.s3);
-
-#endif // M0 > 2
-#if M0 > 3
-
-        acc3.s0 = fma(a3.s0, b0.s0, acc3.s0);
-        acc3.s1 = fma(a3.s0, b0.s1, acc3.s1);
-        acc3.s2 = fma(a3.s0, b0.s2, acc3.s2);
-        acc3.s3 = fma(a3.s0, b0.s3, acc3.s3);
-#endif // M0 > 3
-
-        // Load values from matrix A and matrix B
-        b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-
-        // Multiply and accumulate
-        acc0.s0 = fma(a0.s1, b0.s0, acc0.s0);
-        acc0.s1 = fma(a0.s1, b0.s1, acc0.s1);
-        acc0.s2 = fma(a0.s1, b0.s2, acc0.s2);
-        acc0.s3 = fma(a0.s1, b0.s3, acc0.s3);
-
-#if M0 > 1
-
-        acc1.s0 = fma(a1.s1, b0.s0, acc1.s0);
-        acc1.s1 = fma(a1.s1, b0.s1, acc1.s1);
-        acc1.s2 = fma(a1.s1, b0.s2, acc1.s2);
-        acc1.s3 = fma(a1.s1, b0.s3, acc1.s3);
-
-#endif // M0 > 1
-#if M0 > 2
-
-        acc2.s0 = fma(a2.s1, b0.s0, acc2.s0);
-        acc2.s1 = fma(a2.s1, b0.s1, acc2.s1);
-        acc2.s2 = fma(a2.s1, b0.s2, acc2.s2);
-        acc2.s3 = fma(a2.s1, b0.s3, acc2.s3);
-
-#endif // M0 > 2
-#if M0 > 3
-
-        acc3.s0 = fma(a3.s1, b0.s0, acc3.s0);
-        acc3.s1 = fma(a3.s1, b0.s1, acc3.s1);
-        acc3.s2 = fma(a3.s1, b0.s2, acc3.s2);
-        acc3.s3 = fma(a3.s1, b0.s3, acc3.s3);
-#endif // M0 > 3
-
-        // Load values from matrix A and matrix B
-        b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-
-        // Multiply and accumulate
-        acc0.s0 = fma(a0.s2, b0.s0, acc0.s0);
-        acc0.s1 = fma(a0.s2, b0.s1, acc0.s1);
-        acc0.s2 = fma(a0.s2, b0.s2, acc0.s2);
-        acc0.s3 = fma(a0.s2, b0.s3, acc0.s3);
-
-#if M0 > 1
-
-        acc1.s0 = fma(a1.s2, b0.s0, acc1.s0);
-        acc1.s1 = fma(a1.s2, b0.s1, acc1.s1);
-        acc1.s2 = fma(a1.s2, b0.s2, acc1.s2);
-        acc1.s3 = fma(a1.s2, b0.s3, acc1.s3);
-
-#endif // M0 > 1
-#if M0 > 2
-
-        acc2.s0 = fma(a2.s2, b0.s0, acc2.s0);
-        acc2.s1 = fma(a2.s2, b0.s1, acc2.s1);
-        acc2.s2 = fma(a2.s2, b0.s2, acc2.s2);
-        acc2.s3 = fma(a2.s2, b0.s3, acc2.s3);
-
-#endif // M0 > 2
-#if M0 > 3
-
-        acc3.s0 = fma(a3.s2, b0.s0, acc3.s0);
-        acc3.s1 = fma(a3.s2, b0.s1, acc3.s1);
-        acc3.s2 = fma(a3.s2, b0.s2, acc3.s2);
-        acc3.s3 = fma(a3.s2, b0.s3, acc3.s3);
-#endif // M0 > 3
-
-        // Load values from matrix A and matrix B
-        b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-
-        // Multiply and accumulate
-        acc0.s0 = fma(a0.s3, b0.s0, acc0.s0);
-        acc0.s1 = fma(a0.s3, b0.s1, acc0.s1);
-        acc0.s2 = fma(a0.s3, b0.s2, acc0.s2);
-        acc0.s3 = fma(a0.s3, b0.s3, acc0.s3);
-
-#if M0 > 1
-
-        acc1.s0 = fma(a1.s3, b0.s0, acc1.s0);
-        acc1.s1 = fma(a1.s3, b0.s1, acc1.s1);
-        acc1.s2 = fma(a1.s3, b0.s2, acc1.s2);
-        acc1.s3 = fma(a1.s3, b0.s3, acc1.s3);
-
-#endif // M0 > 1
-#if M0 > 2
-
-        acc2.s0 = fma(a2.s3, b0.s0, acc2.s0);
-        acc2.s1 = fma(a2.s3, b0.s1, acc2.s1);
-        acc2.s2 = fma(a2.s3, b0.s2, acc2.s2);
-        acc2.s3 = fma(a2.s3, b0.s3, acc2.s3);
-
-#endif // M0 > 2
-#if M0 > 3
-
-        acc3.s0 = fma(a3.s3, b0.s0, acc3.s0);
-        acc3.s1 = fma(a3.s3, b0.s1, acc3.s1);
-        acc3.s2 = fma(a3.s3, b0.s2, acc3.s2);
-        acc3.s3 = fma(a3.s3, b0.s3, acc3.s3);
-#endif // M0 > 3
-
-        src_addr.s0 += 4 * sizeof(float);
-    }
-
-    for(; i < (int)K; ++i)
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0));
-#if M0 > 1
-        float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1));
-#endif // M0 > 1
-#if M0 > 2
-        float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2));
-#endif // M0 > 2
-#if M0 > 3
-        float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3));
-#endif // M0 > 3
-#else  // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
-#if M0 > 1
-        float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // M0 > 1
-#if M0 > 2
-        float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // M0 > 2
-#if M0 > 3
-        float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // M0 > 3
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        // Load values from matrix B
-        float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-
-        // Multiply and accumulate
-        acc0.s0 = fma(a0, b0.s0, acc0.s0);
-        acc0.s1 = fma(a0, b0.s1, acc0.s1);
-        acc0.s2 = fma(a0, b0.s2, acc0.s2);
-        acc0.s3 = fma(a0, b0.s3, acc0.s3);
-#if M0 > 1
-        acc1.s0 = fma(a1, b0.s0, acc1.s0);
-        acc1.s1 = fma(a1, b0.s1, acc1.s1);
-        acc1.s2 = fma(a1, b0.s2, acc1.s2);
-        acc1.s3 = fma(a1, b0.s3, acc1.s3);
-#endif // M0 > 1
-#if M0 > 2
-        acc2.s0 = fma(a2, b0.s0, acc2.s0);
-        acc2.s1 = fma(a2, b0.s1, acc2.s1);
-        acc2.s2 = fma(a2, b0.s2, acc2.s2);
-        acc2.s3 = fma(a2, b0.s3, acc2.s3);
-#endif // M0 > 2
-#if M0 > 3
-        acc3.s0 = fma(a3, b0.s0, acc3.s0);
-        acc3.s1 = fma(a3, b0.s1, acc3.s1);
-        acc3.s2 = fma(a3, b0.s2, acc3.s2);
-        acc3.s3 = fma(a3, b0.s3, acc3.s3);
-#endif // M0 > 3
-
-        src_addr.s0 += sizeof(float);
-    }
-
-    int z = get_global_id(2);
-
-    // Compute dst address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0,
-                               PARTIAL_STORE_M0)
-                               * dst_stride_y);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (dst_cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(M0, float, acc, ALPHA);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float));
-
-    LOAD_BLOCK(1, 4, float, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, float, bias, BETA);
-#endif // UNIT_BIAS
-
-    // acc = acc + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(M0, acc, bias0);
-
-#else // defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0,
-                                PARTIAL_STORE_M0)
-                                * src2_stride_y)
-                                + z * src2_stride_z;
-
-    LOAD_BLOCK(M0, 4, float, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(M0, float, bias, BETA);
-#endif // UNIT_BIAS
-
-    // acc = acc + bias
-    ADD_BLOCK(M0, acc, bias);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, float, VEC_SIZE, acc, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store the output block
-    const bool cond_y = get_global_id(1) == 0;
-    const bool cond_x = ((get_global_id(0) + 1) * 4 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(M0, 4, float, acc, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-
-/** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not been reshaped
- *
- * @note This OpenCL kernel works with the 32-bit floating point data type (float) and uses the fma units.
- * This OpenCL kernel is optimized for Bifrost when the number of matrix B columns is less or equal to 1000.
- * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0.
- * @note This kernel processed a fixed number of elements along x: -DN0=2.
- * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
- *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
- *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
- *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
- *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
- *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
- * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
- * @param[in]  dst_step_x                         dst_gx_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_gx_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  src_cross_plane_pad                (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D)
- * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_floating_point_f32_bifrost_1000(IMAGE_DECLARATION(src0),
-                                                      IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                                      IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                                      IMAGE_DECLARATION(dst),
-                                                      uint src0_stride_z,
-                                                      uint src1_stride_z,
-#if defined(BETA)
-                                                      uint src2_stride_z,
-#endif //defined(BETA)
-                                                      uint dst_stride_z
-#if defined(REINTERPRET_INPUT_AS_3D)
-                                                      ,
-                                                      uint src_cross_plane_pad
-#endif // REINTERPRET_INPUT_AS_3D
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                                      ,
-                                                      uint dst_cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                                     )
-{
-    // Requires 2 N0, C vect2, A vect4, B (2 vload2) // to fix for M0 > 1
-    int idx = get_global_id(0) * N0;
-
-    // Compute starting address for matrix A and Matrix B
-    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
-
-    // Update address for the matrix A
-    src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y;
-
-    // Update address for the matrix B
-    src_addr.s1 += idx * sizeof(float);
-
-#if defined(REINTERPRET_INPUT_AS_3D)
-    // Since we load a 2D input tile from 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 row by HEIGHT_GEMM3D
-    uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zin       = min(DEPTH_GEMM3D - 1, zin);
-
-    // Add offset due to the cross plane paddings
-    zin *= (src_cross_plane_pad * src0_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply src0_stride_z by DEPTH_GEMM3D
-    src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;
-
-#else // defined(REINTERPRET_INPUT_AS_3D)
-
-    // Add offset for batched GEMM
-    src_addr.s0 += get_global_id(2) * src0_stride_z;
-
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src_addr.s1 += get_global_id(2) * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    // Initialize accumulators
-    float2 acc0 = 0.0f;
-#if M0 > 1
-    float2 acc1 = 0.0f;
-#endif // M0 > 1
-#if M0 > 2
-    float2 acc2 = 0.0f;
-#endif // M0 > 2
-#if M0 > 3
-    float2 acc3 = 0.0f;
-#endif // M0 > 3
-
-    // A and B src indices get incremented at the same time.
-    int i = 0;
-    for(; i <= ((int)K - 8); i += 8)
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        float8 a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + zin.s0));
-#else  // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        float8 a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0));
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        // Load values from matrix B
-        float2 b0 = vload2(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        float2 b1 = vload2(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        float2 b2 = vload2(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        float2 b3 = vload2(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        float2 b4 = vload2(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        float2 b5 = vload2(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        float2 b6 = vload2(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        float2 b7 = vload2(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-
-        // Multiply and accumulate
-        acc0.s0 = fma(a0.s0, b0.s0, acc0.s0);
-        acc0.s0 = fma(a0.s1, b1.s0, acc0.s0);
-        acc0.s0 = fma(a0.s2, b2.s0, acc0.s0);
-        acc0.s0 = fma(a0.s3, b3.s0, acc0.s0);
-        acc0.s0 = fma(a0.s4, b4.s0, acc0.s0);
-        acc0.s0 = fma(a0.s5, b5.s0, acc0.s0);
-        acc0.s0 = fma(a0.s6, b6.s0, acc0.s0);
-        acc0.s0 = fma(a0.s7, b7.s0, acc0.s0);
-
-        acc0.s1 = fma(a0.s0, b0.s1, acc0.s1);
-        acc0.s1 = fma(a0.s1, b1.s1, acc0.s1);
-        acc0.s1 = fma(a0.s2, b2.s1, acc0.s1);
-        acc0.s1 = fma(a0.s3, b3.s1, acc0.s1);
-        acc0.s1 = fma(a0.s4, b4.s1, acc0.s1);
-        acc0.s1 = fma(a0.s5, b5.s1, acc0.s1);
-        acc0.s1 = fma(a0.s6, b6.s1, acc0.s1);
-        acc0.s1 = fma(a0.s7, b7.s1, acc0.s1);
-
-#if M0 > 1
-#if defined(REINTERPRET_INPUT_AS_3D)
-        a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1));
-#else  // defined(REINTERPRET_INPUT_AS_3D)
-        a0                    = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-        acc1.s0 = fma(a0.s0, b0.s0, acc1.s0);
-        acc1.s0 = fma(a0.s1, b1.s0, acc1.s0);
-        acc1.s0 = fma(a0.s2, b2.s0, acc1.s0);
-        acc1.s0 = fma(a0.s3, b3.s0, acc1.s0);
-        acc1.s0 = fma(a0.s4, b4.s0, acc1.s0);
-        acc1.s0 = fma(a0.s5, b5.s0, acc1.s0);
-        acc1.s0 = fma(a0.s6, b6.s0, acc1.s0);
-        acc1.s0 = fma(a0.s7, b7.s0, acc1.s0);
-
-        acc1.s1 = fma(a0.s0, b0.s1, acc1.s1);
-        acc1.s1 = fma(a0.s1, b1.s1, acc1.s1);
-        acc1.s1 = fma(a0.s2, b2.s1, acc1.s1);
-        acc1.s1 = fma(a0.s3, b3.s1, acc1.s1);
-        acc1.s1 = fma(a0.s4, b4.s1, acc1.s1);
-        acc1.s1 = fma(a0.s5, b5.s1, acc1.s1);
-        acc1.s1 = fma(a0.s6, b6.s1, acc1.s1);
-        acc1.s1 = fma(a0.s7, b7.s1, acc1.s1);
-#endif // M0 > 1
-#if M0 > 2
-#if defined(REINTERPRET_INPUT_AS_3D)
-        a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2));
-#else  // defined(REINTERPRET_INPUT_AS_3D)
-        a0                    = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-        acc2.s0 = fma(a0.s0, b0.s0, acc2.s0);
-        acc2.s0 = fma(a0.s1, b1.s0, acc2.s0);
-        acc2.s0 = fma(a0.s2, b2.s0, acc2.s0);
-        acc2.s0 = fma(a0.s3, b3.s0, acc2.s0);
-        acc2.s0 = fma(a0.s4, b4.s0, acc2.s0);
-        acc2.s0 = fma(a0.s5, b5.s0, acc2.s0);
-        acc2.s0 = fma(a0.s6, b6.s0, acc2.s0);
-        acc2.s0 = fma(a0.s7, b7.s0, acc2.s0);
-
-        acc2.s1 = fma(a0.s0, b0.s1, acc2.s1);
-        acc2.s1 = fma(a0.s1, b1.s1, acc2.s1);
-        acc2.s1 = fma(a0.s2, b2.s1, acc2.s1);
-        acc2.s1 = fma(a0.s3, b3.s1, acc2.s1);
-        acc2.s1 = fma(a0.s4, b4.s1, acc2.s1);
-        acc2.s1 = fma(a0.s5, b5.s1, acc2.s1);
-        acc2.s1 = fma(a0.s6, b6.s1, acc2.s1);
-        acc2.s1 = fma(a0.s7, b7.s1, acc2.s1);
-#endif // M0 > 2
-#if M0 > 3
-#if defined(REINTERPRET_INPUT_AS_3D)
-        a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3));
-#else  // defined(REINTERPRET_INPUT_AS_3D)
-        a0                    = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-        acc3.s0 = fma(a0.s0, b0.s0, acc3.s0);
-        acc3.s0 = fma(a0.s1, b1.s0, acc3.s0);
-        acc3.s0 = fma(a0.s2, b2.s0, acc3.s0);
-        acc3.s0 = fma(a0.s3, b3.s0, acc3.s0);
-        acc3.s0 = fma(a0.s4, b4.s0, acc3.s0);
-        acc3.s0 = fma(a0.s5, b5.s0, acc3.s0);
-        acc3.s0 = fma(a0.s6, b6.s0, acc3.s0);
-        acc3.s0 = fma(a0.s7, b7.s0, acc3.s0);
-
-        acc3.s1 = fma(a0.s0, b0.s1, acc3.s1);
-        acc3.s1 = fma(a0.s1, b1.s1, acc3.s1);
-        acc3.s1 = fma(a0.s2, b2.s1, acc3.s1);
-        acc3.s1 = fma(a0.s3, b3.s1, acc3.s1);
-        acc3.s1 = fma(a0.s4, b4.s1, acc3.s1);
-        acc3.s1 = fma(a0.s5, b5.s1, acc3.s1);
-        acc3.s1 = fma(a0.s6, b6.s1, acc3.s1);
-        acc3.s1 = fma(a0.s7, b7.s1, acc3.s1);
-#endif // M0 > 3
-
-        src_addr.s0 += sizeof(float) * 8;
-    }
-    // float size increment
-    for(; i < (int)K; ++i)
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0));
-#if M0 > 1
-        float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1));
-#endif // M0 > 1
-#if M0 > 2
-        float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2));
-#endif // M0 > 2
-#if M0 > 3
-        float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3));
-#endif // M0 > 3
-#else  // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
-#if M0 > 1
-        float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // M0 > 1
-#if M0 > 2
-        float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // M0 > 2
-#if M0 > 3
-        float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // M0 > 3
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        // Load values from matrix B
-        float2 b0 = vload2(0, (__global float *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-
-        // Multiply and accumulate
-        acc0.s0 = fma(a0, b0.s0, acc0.s0);
-        acc0.s1 = fma(a0, b0.s1, acc0.s1);
-#if M0 > 1
-        acc1.s0 = fma(a1, b0.s0, acc1.s0);
-        acc1.s1 = fma(a1, b0.s1, acc1.s1);
-#endif // M0 > 1
-#if M0 > 2
-        acc2.s0 = fma(a2, b0.s0, acc2.s0);
-        acc2.s1 = fma(a2, b0.s1, acc2.s1);
-#endif // M0 > 2
-#if M0 > 3
-        acc3.s0 = fma(a3, b0.s0, acc3.s0);
-        acc3.s1 = fma(a3, b0.s1, acc3.s1);
-#endif // M0 > 3
-
-        src_addr.s0 += sizeof(float);
-    }
-
-    int z = get_global_id(2);
-
-    // Compute dst address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)2 * sizeof(float)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0,
-                               PARTIAL_STORE_M0)
-                               * dst_stride_y);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (dst_cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(M0, float, acc, ALPHA);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)2 * sizeof(float));
-
-    LOAD_BLOCK(1, 2, float, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, float, bias, BETA);
-#endif // UNIT_BIAS
-
-    // acc = acc + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(M0, acc, bias0);
-
-#else // defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)2 * sizeof(float)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0,
-                                PARTIAL_STORE_M0)
-                                * src2_stride_y)
-                                + z * src2_stride_z;
-
-    LOAD_BLOCK(M0, 2, float, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(M0, float, bias, BETA);
-#endif // UNIT_BIAS
-
-    // acc = acc + bias
-    ADD_BLOCK(M0, acc, bias);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, float, VEC_SIZE, acc, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store the output block
-    const bool cond_y = get_global_id(1) == 0;
-    const bool cond_x = ((get_global_id(0) + 1) * 2 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(M0, 2, float, acc, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-
-#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
-/** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped
- *
- * @note This OpenCL kernel works with the 16-bit floating point data type (half) and accumulating the result in a 32 floating point variable.
- * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0.
- * @note This kernel processed a fixed number of elements along x: -DN0=8.
- * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
- *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
- *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
- *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
- *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
- *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
- * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
- * @param[in]  dst_step_x                         dst_gx_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_gx_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  src_cross_plane_pad                (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D)
- * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_floating_point_f16_bifrost_acc32(IMAGE_DECLARATION(src0),
-                                                       IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                                       IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                                       IMAGE_DECLARATION(dst),
-                                                       uint src0_stride_z,
-                                                       uint src1_stride_z,
-#if defined(BETA)
-                                                       uint src2_stride_z,
-#endif //defined(BETA)
-                                                       uint dst_stride_z
-#if defined(REINTERPRET_INPUT_AS_3D)
-                                                       ,
-                                                       uint src_cross_plane_pad
-#endif // REINTERPRET_INPUT_AS_3D
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                                       ,
-                                                       uint dst_cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                                      )
-{
-    int idx = get_global_id(0) * N0;
-
-    // Compute starting address for matrix A and Matrix B
-    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
-
-    // Update address for the matrix A
-    src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y;
-
-    // Update address for the matrix B
-    src_addr.s1 += idx * sizeof(half);
-
-#if defined(REINTERPRET_INPUT_AS_3D)
-    // Since we load a 2D input tile from 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 row by HEIGHT_GEMM3D
-    uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zin       = min(DEPTH_GEMM3D - 1, zin);
-
-    // Add offset due to the cross plane paddings
-    zin *= (src_cross_plane_pad * src0_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply src0_stride_z by DEPTH_GEMM3D
-    src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;
-
-#else // defined(REINTERPRET_INPUT_AS_3D)
-
-    // Add offset for batched GEMM
-    src_addr.s0 += get_global_id(2) * src0_stride_z;
-
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src_addr.s1 += get_global_id(2) * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    float8 acc0 = 0.0h;
-#if M0 > 1
-    float8 acc1 = 0.0h;
-#endif // M0 > 1
-#if M0 > 2
-    float8 acc2 = 0.0h;
-#endif // M0 > 2
-#if M0 > 3
-    float8 acc3 = 0.0h;
-#endif // M0 > 3
-
-    int i = 0;
-    for(; i <= ((int)K - 4); i += 4)
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        LOAD_BLOCK(M0, 4, half, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s);
-#else // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
-#if M0 > 1
-        half4 a1 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // M0 > 1
-#if M0 > 2
-        half4 a2 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // M0 > 2
-#if M0 > 3
-        half4 a3 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // M0 > 3
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        // Load values from matrix B
-        float8 b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1)));
-        src_addr.s1 += src1_stride_y;
-
-        // Accumulate
-        acc0 = fma(b0, (float8)a0.s0, acc0);
-#if M0 > 1
-        acc1 = fma(b0, (float8)a1.s0, acc1);
-#endif // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (float8)a2.s0, acc2);
-#endif // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (float8)a3.s0, acc3);
-#endif // M0 > 3
-
-        b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1)));
-        src_addr.s1 += src1_stride_y;
-        acc0 = fma(b0, (float8)a0.s1, acc0);
-#if M0 > 1
-        acc1 = fma(b0, (float8)a1.s1, acc1);
-#endif // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (float8)a2.s1, acc2);
-#endif // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (float8)a3.s1, acc3);
-#endif // M0 > 3
-
-        b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1)));
-        src_addr.s1 += src1_stride_y;
-        acc0 = fma(b0, (float8)a0.s2, acc0);
-#if M0 > 1
-        acc1 = fma(b0, (float8)a1.s2, acc1);
-#endif // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (float8)a2.s2, acc2);
-#endif // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (float8)a3.s2, acc3);
-#endif // M0 > 3
-
-        b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1)));
-        src_addr.s1 += src1_stride_y;
-        acc0 = fma(b0, (float8)a0.s3, acc0);
-#if M0 > 1
-        acc1 = fma(b0, (float8)a1.s3, acc1);
-#endif // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (float8)a2.s3, acc2);
-#endif // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (float8)a3.s3, acc3);
-#endif // M0 > 3
-
-        src_addr.s0 += 4 * sizeof(half);
-    }
-
-    for(; i < (int)K; ++i)
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0));
-#if M0 > 1
-        half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1));
-#endif // M0 > 1
-#if M0 > 2
-        half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2));
-#endif // M0 > 2
-#if M0 > 3
-        half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3));
-#endif // M0 > 3
-#else  // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
-#if M0 > 1
-        half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // M0 > 1
-#if M0 > 2
-        half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // M0 > 2
-#if M0 > 3
-        half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // M0 > 3
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        // Load values from matrix B
-        float8 b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1)));
-
-        src_addr += (int2)(sizeof(half), src1_stride_y);
-
-        // Accumulate
-        acc0 = fma(b0, (float8)a0, acc0); // b0 * (half8)a0;
-#if M0 > 1
-        acc1 = fma(b0, (float8)a1, acc1); // b0 * (half8)a1;
-#endif                                    // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (float8)a2, acc2); // b0 * (half8)a2;
-#endif                                    // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (float8)a3, acc3); // b0 * (half8)a3;
-#endif                                    // M0 > 3
-    }
-
-    int z = get_global_id(2);
-
-    // Compute dst address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * dst_stride_y);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (dst_cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(M0, float, acc, ALPHA);
-#endif // defined(ALPHA)
-
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half));
-
-    LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-    float8 bias_f0 = convert_float8(bias0);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, float, bias_f, BETA);
-#endif // UNIT_BIAS
-
-    // acc = acc + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(M0, acc, bias_f0);
-
-#else // defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0,
-                                PARTIAL_STORE_M0)
-                                * src2_stride_y)
-                                + z * src2_stride_z;
-
-    LOAD_BLOCK(M0, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-    float8 bias_f0 = convert_float8(bias0);
-#if M0 > 1
-    float8 bias_f1 = convert_float8(bias1);
-#endif // M0 > 1
-#if M0 > 2
-    float8 bias_f2 = convert_float8(bias2);
-#endif // M0 > 2
-#if M0 > 3
-    float8 bias_f3 = convert_float8(bias3);
-#endif // M0 > 3
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(M0, float, bias_f, BETA);
-#endif // UNIT_BIAS
-
-    // acc = acc + bias
-    ADD_BLOCK(M0, acc, bias_f);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-    half8 acc_h0 = convert_half8(acc0);
-#if M0 > 1
-    half8 acc_h1 = convert_half8(acc1);
-#endif // M0 > 1
-#if M0 > 2
-    half8 acc_h2 = convert_half8(acc2);
-#endif // M0 > 2
-#if M0 > 3
-    half8 acc_h3 = convert_half8(acc3);
-#endif // M0 > 3
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, half, VEC_SIZE, acc_h, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store the output block
-    const bool cond_y = get_global_id(1) == 0;
-    const bool cond_x = ((get_global_id(0) + 1) * 8 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(M0, 8, half, acc_h, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-
-/** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped
- *
- * @note This OpenCL kernel works with the 16-bit floating point data type (half) and uses the fma units.
- * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0.
- * @note This kernel processed a fixed number of elements along x: -DN0=8.
- * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @note The optional alpha's value need to be passed at compile time using -DALPHA
- * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16)
- *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16])
- *
- * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
- *       The activation function is performed after the bias addition
- * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
- *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
- *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
- *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
- *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
- *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
- *
- * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16
- * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
- * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
- * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
- * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
- * @param[in]  src2_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  src2_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
- * @param[in]  src2_step_x                        (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src2_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
- * @param[in]  src2_step_y                        (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
- * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
- * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
- * @param[in]  dst_step_x                         dst_gx_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_gx_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]  src0_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
- * @param[in]  src2_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  src_cross_plane_pad                (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D)
- * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
- */
-__kernel void gemm_mm_floating_point_f16_bifrost(IMAGE_DECLARATION(src0),
-                                                 IMAGE_DECLARATION(src1),
-#if defined(BETA)
-                                                 IMAGE_DECLARATION(src2),
-#endif // defined(BETA)
-                                                 IMAGE_DECLARATION(dst),
-                                                 uint src0_stride_z,
-                                                 uint src1_stride_z,
-#if defined(BETA)
-                                                 uint src2_stride_z,
-#endif //defined(BETA)
-                                                 uint dst_stride_z
-#if defined(REINTERPRET_INPUT_AS_3D)
-                                                 ,
-                                                 uint src_cross_plane_pad
-#endif // REINTERPRET_INPUT_AS_3D
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-                                                 ,
-                                                 uint dst_cross_plane_pad
-#endif // REINTERPRET_OUTPUT_AS_3D
-                                                )
-{
-    int idx = get_global_id(0) * N0;
-
-    // Compute starting address for matrix A and Matrix B
-    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
-
-    // Update address for the matrix A
-    src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y;
-
-    // Update address for the matrix B
-    src_addr.s1 += idx * sizeof(half);
-
-#if defined(REINTERPRET_INPUT_AS_3D)
-    // Since we load a 2D input tile from 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 row by HEIGHT_GEMM3D
-    uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zin       = min(DEPTH_GEMM3D - 1, zin);
-
-    // Add offset due to the cross plane paddings
-    zin *= (src_cross_plane_pad * src0_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply src0_stride_z by DEPTH_GEMM3D
-    src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;
-
-#else // defined(REINTERPRET_INPUT_AS_3D)
-
-    // Add offset for batched GEMM
-    src_addr.s0 += get_global_id(2) * src0_stride_z;
-
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    src_addr.s1 += get_global_id(2) * src1_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    half8 acc0 = 0.0h;
-#if M0 > 1
-    half8 acc1 = 0.0h;
-#endif // M0 > 1
-#if M0 > 2
-    half8 acc2 = 0.0h;
-#endif // M0 > 2
-#if M0 > 3
-    half8 acc3 = 0.0h;
-#endif // M0 > 3
-
-    int i = 0;
-    for(; i <= ((int)K - 4); i += 4)
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        LOAD_BLOCK(M0, 4, half, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s);
-#else // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
-#if M0 > 1
-        half4 a1 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // M0 > 1
-#if M0 > 2
-        half4 a2 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // M0 > 2
-#if M0 > 3
-        half4 a3 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // M0 > 3
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        // Load values from matrix B
-        half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-
-        // Accumulate
-        acc0 = fma(b0, (half8)a0.s0, acc0);
-#if M0 > 1
-        acc1 = fma(b0, (half8)a1.s0, acc1);
-#endif // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (half8)a2.s0, acc2);
-#endif // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (half8)a3.s0, acc3);
-#endif // M0 > 3
-
-        b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        acc0 = fma(b0, (half8)a0.s1, acc0);
-#if M0 > 1
-        acc1 = fma(b0, (half8)a1.s1, acc1);
-#endif // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (half8)a2.s1, acc2);
-#endif // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (half8)a3.s1, acc3);
-#endif // M0 > 3
-
-        b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        acc0 = fma(b0, (half8)a0.s2, acc0);
-#if M0 > 1
-        acc1 = fma(b0, (half8)a1.s2, acc1);
-#endif // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (half8)a2.s2, acc2);
-#endif // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (half8)a3.s2, acc3);
-#endif // M0 > 3
-
-        b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1));
-        src_addr.s1 += src1_stride_y;
-        acc0 = fma(b0, (half8)a0.s3, acc0);
-#if M0 > 1
-        acc1 = fma(b0, (half8)a1.s3, acc1);
-#endif // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (half8)a2.s3, acc2);
-#endif // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (half8)a3.s3, acc3);
-#endif // M0 > 3
-
-        src_addr.s0 += 4 * sizeof(half);
-    }
-
-    for(; i < (int)K; ++i)
-    {
-#if defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0));
-#if M0 > 1
-        half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1));
-#endif // M0 > 1
-#if M0 > 2
-        half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2));
-#endif // M0 > 2
-#if M0 > 3
-        half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3));
-#endif // M0 > 3
-#else  // defined(REINTERPRET_INPUT_AS_3D)
-        // Load values from matrix A
-        half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));
-#if M0 > 1
-        half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));
-#endif // M0 > 1
-#if M0 > 2
-        half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));
-#endif // M0 > 2
-#if M0 > 3
-        half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));
-#endif // M0 > 3
-#endif // defined(REINTERPRET_INPUT_AS_3D)
-
-        // Load values from matrix B
-        half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1));
-
-        src_addr += (int2)(sizeof(half), src1_stride_y);
-
-        // Accumulate
-        acc0 = fma(b0, (half8)a0, acc0); // b0 * (half8)a0;
-#if M0 > 1
-        acc1 = fma(b0, (half8)a1, acc1); // b0 * (half8)a1;
-#endif                                   // M0 > 1
-#if M0 > 2
-        acc2 = fma(b0, (half8)a2, acc2); // b0 * (half8)a2;
-#endif                                   // M0 > 2
-#if M0 > 3
-        acc3 = fma(b0, (half8)a3, acc3); // b0 * (half8)a3;
-#endif                                   // M0 > 3
-    }
-
-    int z = get_global_id(2);
-
-    // Compute dst address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * dst_stride_y);
-
-    uint4 zout = 0;
-
-#if defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension
-    // in order to take into account the presence of possible cross plane paddings
-    //
-    //  |                  |
-    //  |      plane0      |
-    //  |                  |
-    //  |__________________|
-    //  |******************|
-    //  |  cross_plane_pad |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |                  |
-    //  |__________________|
-
-    // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D
-    zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D;
-    zout = min(DEPTH_GEMM3D - 1, zout);
-
-    // Add offset due to the cross plane paddings
-    zout *= (dst_cross_plane_pad * dst_stride_y);
-
-    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
-    // multiply dst_stride_z by DEPTH_GEMM3D
-    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
-#else  // defined(REINTERPRET_OUTPUT_AS_3D)
-    // Add offset for batched GEMM
-    dst_addr += z * dst_stride_z;
-#endif // defined(REINTERPRET_OUTPUT_AS_3D)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    SCALE_BLOCK(M0, half, acc, ALPHA);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-    REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0);
-
-#if defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half));
-
-    LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(1, half, bias, BETA);
-#endif // UNIT_BIAS
-
-    // acc = acc + bias[broadcasted]
-    ADD_BLOCK_BROADCAST(M0, acc, bias0);
-
-#else // defined(BROADCAST_BIAS)
-    __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0,
-                                PARTIAL_STORE_M0)
-                                * src2_stride_y)
-                                + z * src2_stride_z;
-
-    LOAD_BLOCK(M0, 8, half, bias, src2_addr, 0, src2_stride_y, zero);
-
-#ifndef UNIT_BETA
-    SCALE_BLOCK(M0, half, bias, BETA);
-#endif // UNIT_BIAS
-
-    // acc = acc + bias
-    ADD_BLOCK(M0, acc, bias);
-
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(ACTIVATION_TYPE)
-    ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, half, VEC_SIZE, acc, A_VAL, B_VAL);
-#endif // defined(ACTIVATION_TYPE)
-
-    // Store the output block
-    const bool cond_y = get_global_id(1) == 0;
-    const bool cond_x = ((get_global_id(0) + 1) * 8 >= N);
-    STORE_BLOCK_BOUNDARY_AWARE(M0, 8, half, acc, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
-}
-#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
-
-#endif // defined(N) && defined(K) && defined(M0) && defined(N0) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
\ No newline at end of file
diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp
index 4af4226..9d524f9 100644
--- a/src/gpu/cl/ClKernelLibrary.cpp
+++ b/src/gpu/cl/ClKernelLibrary.cpp
@@ -271,16 +271,6 @@
     { "gemm_ma_f32", "common/gemm.cl" },
     { "gemm_mv", "common/gemv.cl" },
     { "gemm_mv_quantized", "common/gemv.cl" },
-    { "gemm_mm_interleaved_transposed_f16", "common/gemm_v1.cl" },
-    { "gemm_mm_interleaved_transposed_f16_acc32", "common/gemm_v1.cl" },
-    { "gemm_mm_interleaved_transposed_f16_bifrost", "common/gemm_v1.cl" },
-    { "gemm_mm_interleaved_transposed_f32", "common/gemm_v1.cl" },
-    { "gemm_mm_interleaved_transposed_f32_bifrost", "common/gemm_v1.cl" },
-    { "gemm_mm_floating_point", "common/gemm_v1.cl" },
-    { "gemm_mm_floating_point_f16_bifrost", "common/gemm_v1.cl" },
-    { "gemm_mm_floating_point_f16_bifrost_acc32", "common/gemm_v1.cl" },
-    { "gemm_mm_floating_point_f32_bifrost", "common/gemm_v1.cl" },
-    { "gemm_mm_floating_point_f32_bifrost_1000", "common/gemm_v1.cl" },
     { "gemm_mm_native", "common/gemm.cl" },
     { "gemm_mm_reshaped_lhs_nt_rhs_t", "common/gemm.cl" },
     { "gemm_mm_reshaped_lhs_nt_rhs_t_texture", "common/gemm.cl" },
@@ -591,10 +581,6 @@
 #include "./cl_kernels/common/gemm.clembed"
     },
     {
-        "common/gemm_v1.cl",
-#include "./cl_kernels/common/gemm_v1.clembed"
-    },
-    {
         "common/gemmlowp.cl",
 #include "./cl_kernels/common/gemmlowp.clembed"
     },
diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.cpp
deleted file mode 100644
index 4e934f0..0000000
--- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.cpp
+++ /dev/null
@@ -1,538 +0,0 @@
-/*
- * Copyright (c) 2017-2021 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 "src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.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/Helpers.h"
-#include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Utils.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "src/core/AccessWindowStatic.h"
-#include "src/core/CL/CLValidate.h"
-#include "src/core/helpers/AutoConfiguration.h"
-#include "src/core/helpers/WindowHelpers.h"
-#include "src/core/utils/helpers/float_ops.h"
-#include "support/Cast.h"
-#include "support/StringSupport.h"
-
-namespace arm_compute
-{
-namespace opencl
-{
-namespace kernels
-{
-namespace
-{
-using ElementsProcessed = Steps;
-
-inline Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float beta,
-                                 bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, bool fp_mixed_precision)
-{
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
-    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src0);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((fp_mixed_precision && (src0->data_type() != DataType::F16)), "Mixed precision floating point is supported only for F16 data");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the matrix B must be <= 3");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_interleaved_transposed && reshape_info.reinterpret_input_as_3d(), "The input tensor cannot be reinterpreted as 3D if is_interleaved_transposed is true");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 2 && reshape_info.reinterpret_input_as_3d(), "The src1 tensor cannot have more than 2 dimensions if src0 has to be reinterpreted as 3D");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((reshape_info.reinterpret_input_as_3d() || reshape_info.depth_output_gemm3d() != 0) && (src2 != nullptr)
-                                    && (!reshape_info.broadcast_bias()),
-                                    "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D");
-
-    if(!is_interleaved_transposed)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != src1->dimension(1));
-
-        if(src2 != nullptr && !(helpers::float_ops::is_zero(beta)))
-        {
-            const unsigned int m         = reshape_info.reinterpret_input_as_3d() ? src0->dimension(1) * src0->dimension(2) : src0->dimension(1);
-            const unsigned int n         = src1->dimension(0);
-            const unsigned int src2_dim0 = src2->dimension(0);
-            const unsigned int src2_dim1 = src2->dimension(1);
-
-            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src2, src1);
-            if(reshape_info.broadcast_bias())
-            {
-                ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
-            }
-            else
-            {
-                ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim0 != n || src2_dim1 != m), "Incorrect dimension of bias matrix");
-            }
-        }
-    }
-    else
-    {
-        GEMMRHSMatrixInfo rhs_info;
-        GEMMLHSMatrixInfo lhs_info;
-        const auto        m                         = static_cast<unsigned int>(reshape_info.m());
-        const auto        n                         = static_cast<unsigned int>(reshape_info.n());
-        const int         k                         = reshape_info.k();
-        const int         mult_transpose1xW_width   = reshape_info.mult_transpose1xW_width();
-        const int         mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
-        rhs_info.n0                                 = max_cl_vector_width / src1->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                          = true;
-
-        TensorShape tensor_shape0{ src0->tensor_shape() };
-        tensor_shape0.set(0, k);
-        tensor_shape0.set(1, m);
-
-        TensorShape tensor_shape1{ src1->tensor_shape() };
-        tensor_shape1.set(0, n);
-        tensor_shape1.set(1, k);
-
-        const TensorInfo tensor_info0 = src0->clone()->set_tensor_shape(tensor_shape0);
-        const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1);
-
-        const TensorInfo tensor_info_reshaped0 = src0->clone()->set_tensor_shape(misc::shape_calculator::compute_lhs_reshaped_shape(tensor_info0, lhs_info));
-        const TensorInfo tensor_info_reshaped1 = src1->clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(tensor_info1, rhs_info));
-
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src0, &tensor_info_reshaped0);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src1, &tensor_info_reshaped1);
-
-        if(src2 != nullptr && !(helpers::float_ops::is_zero(beta)))
-        {
-            const unsigned int src2_dim0 = src2->dimension(0);
-            const unsigned int src2_dim1 = src2->dimension(1);
-
-            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src2, src1);
-            if(reshape_info.broadcast_bias())
-            {
-                ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
-            }
-            else
-            {
-                ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim0 != n || src2_dim1 != m), "Incorrect dimension of bias matrix");
-            }
-        }
-    }
-
-    if(dst->total_size() != 0)
-    {
-        const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, is_interleaved_transposed, reshape_info));
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst);
-    }
-
-    return Status{};
-}
-
-inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst,
-                                                               float beta, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target,
-                                                               ElementsProcessed &num_elements_processed)
-{
-    ARM_COMPUTE_UNUSED(beta);
-    bool   window_changed = false;
-    Window win{};
-    Window win_out{};
-
-    const DataType data_type                           = src0->data_type();
-    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_input_as_3d             = reshape_info.reinterpret_input_as_3d();
-    bool           reinterpret_output_as_3d            = (reshape_info.depth_output_gemm3d() != 0);
-
-    // In case both input and dst have to be reinterpreted as 3D tensors,
-    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
-    if(reinterpret_input_as_3d == reinterpret_output_as_3d)
-    {
-        reinterpret_input_as_3d  = false;
-        reinterpret_output_as_3d = false;
-    }
-
-    // dst tensor auto inizialitation if not yet initialized
-    auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, is_interleaved_transposed, reshape_info)));
-
-    TensorInfo tmp_info(*dst);
-
-    if(reinterpret_output_as_3d)
-    {
-        // Since the dst 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(dst->tensor_shape());
-        tmp_shape.collapse(2U, 1U);
-        tmp_info.set_tensor_shape(tmp_shape);
-    }
-
-    if(is_interleaved_transposed)
-    {
-        // reinterpret_input_as_3d is not supported if is_interleaved_transposed is set
-        ARM_COMPUTE_ERROR_ON(reshape_info.reinterpret_input_as_3d());
-
-        // Configure kernel window
-        num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type);
-        num_elems_processed_per_iteration_y = 4;
-
-        win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-        if(src2 != nullptr)
-        {
-            const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
-
-            const int bias_processed_per_iteration_y = reshape_info.broadcast_bias() ? 1 : num_elems_processed_per_iteration_y;
-
-            AccessWindowStatic src2_access(src2, 0, 0,
-                                           ceil_to_multiple(src2->dimension(0), bias_processed_per_iteration_x),
-                                           ceil_to_multiple(src2->dimension(1), bias_processed_per_iteration_y));
-
-            window_changed = update_window_and_padding(win, src2_access); // window used by the execute_window_loop
-        }
-    }
-    else // The input tensors have not been reshaped
-    {
-        // Special case for 1xN, 2xN, 3xN and 4xN src0 tensor. num_elems_processed_per_iteration_x is set up for the default case.
-        num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type);
-        num_elems_processed_per_iteration_y = std::min(static_cast<int>(dst->dimension(1)), 4);
-
-        // Create kernels according to the architecture, data type and input size.
-        GPUTarget arch_target = get_arch_from_target(gpu_target);
-        if(arch_target == GPUTarget::BIFROST && data_type == DataType::F32)
-        {
-            num_elems_processed_per_iteration_x = (src1->dimension(0) <= 1000 && src0->num_dimensions() == 1) ? 2 : 4;
-        }
-
-        // Configure window
-        win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-        win_out = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-        AccessWindowStatic src0_access(src0, 0, 0, src0->dimension(0), src0->dimension(1));
-        AccessWindowStatic src1_access(src1, 0, 0, ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x), src1->dimension(1));
-        AccessWindowStatic dst_access(dst, 0, 0,
-                                      dst->dimension(0),
-                                      dst->dimension(1));
-
-        if(src2 != nullptr)
-        {
-            const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
-
-            AccessWindowStatic src2_access(src2, 0, 0,
-                                           ceil_to_multiple(src2->dimension(0), bias_processed_per_iteration_x),
-                                           src2->dimension(1));
-
-            window_changed = update_window_and_padding(win, src0_access, src1_access, src2_access) || // window used by the execute_window_loop
-                             update_window_and_padding(win_out, dst_access);                          // window used to update the padding requirements of dst tensor
-        }
-        else
-        {
-            window_changed = update_window_and_padding(win, src0_access, src1_access) || // window used by the execute_window_loop
-                             update_window_and_padding(win_out, dst_access);             // window used to update the padding requirements of dst tensor
-        }
-    }
-
-    // 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>(dst->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
-
-ClGemmMatrixMultiplyKernel::ClGemmMatrixMultiplyKernel()
-{
-    _type = CLKernelType::GEMM;
-}
-
-void ClGemmMatrixMultiplyKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha,
-                                           float beta,
-                                           bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, bool fp_mixed_precision, const ActivationLayerInfo &activation_info)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
-
-    // Perform validate step
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, beta,
-                                                  is_interleaved_transposed, reshape_info, fp_mixed_precision));
-
-    auto padding_info = is_interleaved_transposed ? get_padding_info({ src0, src1, dst }) : get_padding_info({ src0, dst });
-
-    _reinterpret_input_as_3d  = reshape_info.reinterpret_input_as_3d();
-    _reinterpret_output_as_3d = (reshape_info.depth_output_gemm3d() != 0);
-    _add_bias                 = src2 != nullptr;
-
-    // In case both input and dst have to be reinterpreted as 3D tensors,
-    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
-    if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
-    {
-        _reinterpret_input_as_3d  = false;
-        _reinterpret_output_as_3d = false;
-    }
-
-    // Check if we need to slide the matrix B
-    const unsigned int num_dimensions_src0 = _reinterpret_input_as_3d ? src0->num_dimensions() - 1 : src0->num_dimensions();
-
-    _slide_matrix_b = (src1->num_dimensions() >= num_dimensions_src0);
-
-    const DataType data_type = src0->data_type();
-
-    // Get target architecture
-    GPUTarget gpu_target = get_target();
-
-    ElementsProcessed num_elements_processed{};
-
-    // Configure kernel window
-    auto win_config = validate_and_configure_window(src0, src1, src2, dst, beta, is_interleaved_transposed, reshape_info,
-                                                    gpu_target, num_elements_processed);
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
-
-    // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true, both will be turned off (false)
-    // in which case we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
-    // This means that the actual m used by the kernel is given by dst->dimension(1)
-    const unsigned int internal_m = _reinterpret_output_as_3d ? dst->dimension(1) * dst->dimension(2) : dst->dimension(1);
-    const unsigned int n          = dst->dimension(0);
-
-    const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(1) : src0->dimension(1);
-    const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(2) : src0->dimension(2);
-
-    const unsigned int m0 = num_elements_processed.y();
-    const unsigned int n0 = num_elements_processed.x();
-
-    // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding.
-    const unsigned int partial_store_m0 = internal_m % m0;
-    const unsigned int partial_store_n0 = n % n0;
-
-    // Create build options
-    CLBuildOptions build_opts;
-
-    build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
-    build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
-    build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
-    build_opts.add_option_if(reshape_info.broadcast_bias(), "-DBROADCAST_BIAS");
-    build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
-    build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
-    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d));
-    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d));
-    build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
-    build_opts.add_option_if(activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(activation_info.activation())));
-    build_opts.add_option_if(activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(activation_info.a()));
-    build_opts.add_option_if(activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(activation_info.b()));
-    build_opts.add_option("-DIN1_DIM_X=" + support::cpp11::to_string(src1->dimension(0)));
-
-    const bool is_bifrost = get_arch_from_target(gpu_target) == GPUTarget::BIFROST;
-
-    std::string kernel_name;
-    if(is_interleaved_transposed)
-    {
-        const int mult_transpose1xW_width   = reshape_info.mult_transpose1xW_width();
-        const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
-
-        build_opts.add_option("-DM=" + support::cpp11::to_string(internal_m));
-        build_opts.add_option("-DN=" + support::cpp11::to_string(n));
-        build_opts.add_option("-DK=" + support::cpp11::to_string(src1->dimension(0) / (n0 * mult_transpose1xW_width)));
-        build_opts.add_option("-DH0=" + support::cpp11::to_string(mult_transpose1xW_width));
-        build_opts.add_option("-DV0=" + support::cpp11::to_string(mult_interleave4x4_height));
-        build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
-        build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
-
-        if(is_data_type_float(data_type) && is_bifrost)
-        {
-            kernel_name = "gemm_mm_interleaved_transposed_" + lower_string(string_from_data_type(data_type)) + "_bifrost";
-        }
-        else
-        {
-            kernel_name = "gemm_mm_interleaved_transposed_" + lower_string(string_from_data_type(data_type));
-            if(fp_mixed_precision && data_type == DataType::F16)
-            {
-                // currently wider accumulator is only supported for fp16 kernels.
-                kernel_name += "_acc32";
-            }
-        }
-    }
-    else // The input tensors have not been reshaped
-    {
-        build_opts.add_option("-DN=" + support::cpp11::to_string(n));
-        build_opts.add_option("-DK=" + support::cpp11::to_string(src0->dimension(0)));
-        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
-        build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
-        build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
-        build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
-        build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
-
-        // Create kernels according to the architecture, data type and input size.
-        if(is_data_type_float(data_type) && is_bifrost)
-        {
-            kernel_name = "gemm_mm_floating_point";
-
-            if(src0->num_dimensions() != 1)
-            {
-                kernel_name += "_" + lower_string(string_from_data_type(data_type)) + "_bifrost";
-                if(fp_mixed_precision && data_type == DataType::F16)
-                {
-                    // currently wider accumulator is only supported for fp16 kernels.
-                    kernel_name += "_acc32";
-                }
-            }
-            else if(src1->dimension(0) <= 1000 && data_type == DataType::F32)
-            {
-                // The first kernel is optimized for the case of 1000 or less dst elements (e.g. FC8 of AlexNet and VGG-16, and
-                // FC1 of Inception v3). The second kernel is optimized for the case of greater than 1000 dst elements (e.g.
-                // FC6 and FC7 of AlexNet and VGG-16).
-                kernel_name += "_" + lower_string(string_from_data_type(data_type)) + "_bifrost_1000";
-            }
-
-            // The work-group size equal to the Bifrost quad size has been proved to be optimal for these kernels
-            // via exhaustive autotuning over a range of representative layer configurations.
-            set_lws_hint(cl::NDRange(4));
-        }
-        else // (MIDGARD and F32) or (F16)
-        {
-            kernel_name = "gemm_mm_floating_point";
-        }
-    }
-    // Create kernel
-    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
-
-    // Set config_id for enabling LWS tuning
-    _config_id = "gemm_";
-    _config_id += (is_interleaved_transposed ? "reshaped_" : "");
-    _config_id += (_add_bias ? "add_bias_" : "");
-    _config_id += (reshape_info.broadcast_bias() ? "broadcast_bias_" : "");
-    _config_id += (fp_mixed_precision ? "fp_mixed_" : "");
-    _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
-    _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
-    _config_id += lower_string(string_from_data_type(src0->data_type()));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(dst->dimension(1));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(dst->dimension(0));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(dst->dimension(2));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(dst->dimension(3));
-    _config_id += "_";
-    _config_id += (is_interleaved_transposed ? support::cpp11::to_string(src1->dimension(0)) : support::cpp11::to_string(src1->dimension(1)));
-
-    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
-}
-
-Status ClGemmMatrixMultiplyKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta,
-                                            bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target, bool fp_mixed_precision, const ActivationLayerInfo &activation_info)
-{
-    // Note: num_elements_processed will be set in validate_and_configure_window()
-    ElementsProcessed num_elements_processed{};
-    ARM_COMPUTE_UNUSED(alpha);
-    ARM_COMPUTE_UNUSED(activation_info);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, beta, is_interleaved_transposed, reshape_info, fp_mixed_precision));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(),
-                                                              src1->clone().get(),
-                                                              (src2 != nullptr) ? src2->clone().get() : nullptr,
-                                                              dst->clone().get(),
-                                                              beta,
-                                                              is_interleaved_transposed,
-                                                              reshape_info,
-                                                              gpu_target,
-                                                              num_elements_processed)
-                                .first);
-
-    return Status{};
-}
-
-void ClGemmMatrixMultiplyKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
-{
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
-
-    const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
-    const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
-    const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
-    auto       dst  = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
-
-    ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
-    ARM_COMPUTE_ERROR_ON(_add_bias && src2 == nullptr);
-
-    if(src1->info()->num_dimensions() < 3)
-    {
-        // The stride_z for matrix B must be zero if we do not slice
-        ARM_COMPUTE_ERROR_ON(src1->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));
-
-    const unsigned int num_arguments_bias = _add_bias ? num_arguments_per_2D_tensor() + 1 : 0;
-
-    if(_reinterpret_input_as_3d)
-    {
-        // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
-        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3 + num_arguments_bias;
-        const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->info()->padding().bottom;
-        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
-    }
-
-    if(_reinterpret_output_as_3d)
-    {
-        // Pass bottom paddings to the kernel if the dst has to be reinterpreted as 3D tensor
-        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0) + num_arguments_bias;
-        const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->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, src0, slice);
-        add_2D_tensor_argument(idx, src1, slice_b);
-        if(_add_bias)
-        {
-            add_2D_tensor_argument(idx, src2, slice);
-        }
-        add_2D_tensor_argument(idx, dst, slice);
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
-        if(_add_bias)
-        {
-            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[2]));
-        }
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
-        enqueue(queue, *this, slice, lws_hint());
-    }
-    while(window.slide_window_slice_3D(slice));
-}
-} // namespace kernels
-} // namespace opencl
-} // namespace arm_compute
diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.h b/src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.h
deleted file mode 100644
index c16e327..0000000
--- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.h
+++ /dev/null
@@ -1,88 +0,0 @@
-/*
- * Copyright (c) 2017-2021 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_CL_GEMM_MATRIXMULTIPLY_KERNEL_H
-#define ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_KERNEL_H
-
-#include "src/core/common/Macros.h"
-#include "src/gpu/cl/ClCompileContext.h"
-#include "src/gpu/cl/IClKernel.h"
-
-namespace arm_compute
-{
-namespace opencl
-{
-namespace kernels
-{
-/** OpenCL kernel to multiply two input matrices "A" and "B" and add a martix "C" if provided. All elements of the output matrix will be multiplied by alpha. In case matrix C is passed, it will be added to the previous result.
- *  For the matrix C, the broadcast addition is supported if the flag "broadcast_bias" is set in the GEMMReshapeInfo object
- *
- * @note If the input tensors @p src0 and @p src1 have been reshaped respectively with @ref ClGemmReshapeLhsMatrixKernel" and @ref ClGemmReshapeRhsMatrixKernel,
- *       the flag @p is_interleaved_transposed must be set to true
- *
- * @attention @p src1 tensor must have at least 2 dimensions (matrix)
- */
-class ClGemmMatrixMultiplyKernel : public IClKernel
-{
-public:
-    ClGemmMatrixMultiplyKernel();
-    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClGemmMatrixMultiplyKernel);
-    /** Initialise the kernel's input, output and alpha
-     *
-     * @param[in]  compile_context           The compile context to be used.
-     * @param[in]  src0                      Input tensor containing the Matrix A. Data types supported: F16/F32
-     * @param[in]  src1                      Input tensor containing the Matrix B. Data type supported: same as @p src0
-     * @param[in]  src2                      Input tensor containing the Matrix C (bias). Can be nullptr. Data type supported: same as @p src0
-     * @param[out] dst                       Output tensor to store the result of matrix multiplication. Data type supported: same as @p src0
-     * @param[in]  alpha                     Weight of the matrix product
-     * @param[in]  beta                      (Optional) Weight of vector C. Default value is 0. Only beta = 1 is currently supported.
-     * @param[in]  is_interleaved_transposed (Optional) True if input0 and input1 have been reshaped respectively using @ref ClGemmReshapeLhsMatrixKernel and @ref ClGemmReshapeRhsMatrixKernel
-     * @param[in]  reshape_info              (Optional) GEMM reshape info. If is_interleaved_transposed = true, this object must contain the information to understand how the matrix A and matrix B have been reshaped
-     * @param[in]  fp_mixed_precision        (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy
-     * @param[in]  activation_info           (Optional) Activation to apply after the matrix multiplication
-     *
-     */
-    void configure(const ClCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta = 0.f,
-                   bool is_interleaved_transposed = true, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo(), bool fp_mixed_precision = false, const ActivationLayerInfo &activation_info = ActivationLayerInfo());
-    /** Static function to check if given info will lead to a valid configuration
-     *
-     * Similar to @ref ClGemmMatrixMultiplyKernel::configure()
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta,
-                           bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target, bool fp_mixed_precision = false, const ActivationLayerInfo &activation_info = ActivationLayerInfo());
-
-    // Inherited methods overridden:
-    void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
-
-public:
-    bool _slide_matrix_b{ true };
-    bool _reinterpret_input_as_3d{ false };
-    bool _reinterpret_output_as_3d{ false };
-    bool _add_bias{ false };
-};
-} // namespace kernels
-} // namespace opencl
-} // namespace arm_compute
-#endif /* ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_KERNEL_H */
diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
index 448d353..6c872fd 100644
--- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
+++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
@@ -55,7 +55,7 @@
 {
     ARM_COMPUTE_UNUSED(alpha);
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F32, DataType::F16);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h
index 26dec91..89837cc 100644
--- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h
+++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h
@@ -44,7 +44,7 @@
     /** Initialise the kernel's input and dst.
      *
      * @param[in]  compile_context The compile context to be used.
-     * @param[in]  src0            Input tensor for the LHS matrix. Data type supported: F32. The number of dimensions for the LHS matrix must be less or equal than 4.
+     * @param[in]  src0            Input tensor for the LHS matrix. Data type supported: F32/F16. The number of dimensions for the LHS matrix must be less or equal than 4.
      * @param[in]  src1            Input tensor for the RHS matrix. Data type supported: same as @p src0. The number of dimensions for the RHS matrix must be less or equal than 3.
      * @param[in]  src2            Input tensor containing the bias matrix. Data type supported: same as @p src0.
      * @param[out] dst             dst tensor info. Data type supported: same as @p src0
diff --git a/src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeBifrost.cpp b/src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeBifrost.cpp
index b9eac24..d74c7fa 100644
--- a/src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeBifrost.cpp
+++ b/src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeBifrost.cpp
@@ -101,7 +101,7 @@
     }
     else
     {
-        return configure_lhs_rhs_info(m, n, 5, 4, 2, 1, 1, false, false, false, false);
+        return configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 1, false, false, false, false);
     }
 }
 
diff --git a/src/gpu/cl/operators/ClFullyConnected.h b/src/gpu/cl/operators/ClFullyConnected.h
index dc5f9e5..b5ac70c 100644
--- a/src/gpu/cl/operators/ClFullyConnected.h
+++ b/src/gpu/cl/operators/ClFullyConnected.h
@@ -46,7 +46,7 @@
  *
  *  -# @ref opencl::kernels::ClIm2ColKernel (called when the input comes from a convolutional layer)
  *  -# @ref CLTranspose (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
- *  -# @ref opencl::kernels::ClGemmMatrixMultiplyKernel or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
+ *  -# @ref opencl::ClGemm or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
  *
  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
  */
diff --git a/src/gpu/cl/operators/ClGemm.cpp b/src/gpu/cl/operators/ClGemm.cpp
index 4cd5237..d2d0f8f 100644
--- a/src/gpu/cl/operators/ClGemm.cpp
+++ b/src/gpu/cl/operators/ClGemm.cpp
@@ -64,27 +64,14 @@
 {
 inline bool validate_gemm_kernel(CLGEMMKernelType kernel_type)
 {
-    switch(kernel_type)
-    {
-        case CLGEMMKernelType::NATIVE_V1:
-        case CLGEMMKernelType::RESHAPED_ONLY_RHS:
-        case CLGEMMKernelType::RESHAPED_V1:
-        case CLGEMMKernelType::RESHAPED:
-        {
-            return true;
-        }
-        default:
-        {
-            return false;
-        }
-    }
+    return kernel_type == CLGEMMKernelType::NATIVE? false : true;
 }
 //Automatically select between mlgo (prioritized) and default heuristics for gemm kernel type
 inline CLGEMMKernelType auto_select_gemm_kernel(auto_heuristics::CommonQuery query, bool reshape_b_only_on_first_run, bool constant_weights)
 {
     if(!constant_weights)
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
 
     auto gemm_kernel = auto_heuristics::select_mlgo_gemm_kernel(query, reshape_b_only_on_first_run);
@@ -198,97 +185,54 @@
 } // namespace
 
 ClGemm::ClGemm()
-    : _mm_kernel(std::make_unique<ClGemmMatrixMultiplyKernel>()),
-      _reshape_lhs_kernel(std::make_unique<ClGemmReshapeLhsMatrixKernel>()),
+    : _reshape_lhs_kernel(std::make_unique<ClGemmReshapeLhsMatrixKernel>()),
       _reshape_rhs_kernel(std::make_unique<ClGemmReshapeRhsMatrixKernel>()),
+      _mm_native_kernel(std::make_unique<ClGemmMatrixMultiplyNativeKernel>()),
       _mm_reshaped_kernel(std::make_unique<ClGemmMatrixMultiplyReshapedKernel>()),
       _mm_reshaped_only_rhs_kernel(std::make_unique<ClGemmMatrixMultiplyReshapedOnlyRhsKernel>()),
       _mm_reshaped_only_rhs_fallback_kernel(std::make_unique<ClGemmMatrixMultiplyReshapedOnlyRhsKernel>()),
       _tmp_a(),
       _tmp_b(),
       _reshape_b_only_on_first_run(false),
-      _gemm_kernel_type(CLGEMMKernelType::NATIVE_V1),
+      _gemm_kernel_type(CLGEMMKernelType::NATIVE),
       _is_prepared(false),
       _aux_mem(AuxTensorIdx::Count)
 {
 }
 
-void ClGemm::configure_native_v1(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta,
-                                 const GEMMInfo &gemm_info)
+void ClGemm::configure_native(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta,
+                              const GEMMInfo &gemm_info)
 {
-    const unsigned int m          = gemm_info.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 GPUTarget    gpu_target = CLScheduler::get().target();
+    DataType           data_type               = a->data_type();
+    bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
+    const unsigned int m                       = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
+    const unsigned int n                       = b->dimension(0);
+    const unsigned int k                       = a->dimension(0);
+    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();
+    const GPUTarget    gpu_target              = CLScheduler::get().target();
+    bool               broadcast_bias          = gemm_info.broadcast_bias();
+
+    GEMMKernelInfo kernel_info;
+    kernel_info.m                       = m;
+    kernel_info.n                       = n;
+    kernel_info.k                       = k;
+    kernel_info.depth_output_gemm3d     = depth_output_gemm3d;
+    kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d;
+    kernel_info.broadcast_bias          = broadcast_bias;
+    kernel_info.activation_info         = gemm_info.activation_info();
 
     // Set the target for the kernels
-    _mm_kernel->set_target(gpu_target);
+    _mm_native_kernel->set_target(gpu_target);
 
-    GEMMReshapeInfo reshape_info(m, n, k, 1, 1, gemm_info.depth_output_gemm3d(), gemm_info.reinterpret_input_as_3d(), gemm_info.broadcast_bias());
+    auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size });
 
     // Configure and tune matrix multiply kernel
-    _mm_kernel->configure(compile_context, a, b, c, output, alpha, beta, false, reshape_info, gemm_info.fp_mixed_precision(), gemm_info.activation_info());
-
-    // Tune kernel statically
-    CLScheduler::get().tune_kernel_static(*_mm_kernel);
+    _mm_native_kernel->configure(compile_context, a, b, c, output, alpha, beta, config.lhs_info, config.rhs_info, kernel_info);
 }
 
-void ClGemm::configure_reshaped_v1(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta,
-                                   const GEMMInfo &gemm_info)
-{
-    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 int          depth_output_gemm3d       = gemm_info.depth_output_gemm3d();
-    const GPUTarget    gpu_target                = CLScheduler::get().target();
-    int                mult_transpose1xW_width   = 1;
-    int                mult_interleave4x4_height = 1;
-
-    // Set the target for the kernels
-    _reshape_lhs_kernel->set_target(gpu_target);
-    _mm_kernel->set_target(gpu_target);
-
-    if(get_arch_from_target(gpu_target) == GPUTarget::BIFROST)
-    {
-        mult_transpose1xW_width   = 4;
-        mult_interleave4x4_height = 2;
-    }
-
-    GEMMRHSMatrixInfo rhs_info;
-    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;
-
-    GEMMLHSMatrixInfo lhs_info;
-    lhs_info.m0         = 4;
-    lhs_info.k0         = 4;
-    lhs_info.v0         = mult_interleave4x4_height;
-    lhs_info.interleave = true;
-    lhs_info.transpose  = true;
-
-    GEMMReshapeInfo reshape_info(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d, false, gemm_info.broadcast_bias());
-
-    // Configure interleave kernel
-    _reshape_lhs_kernel->configure(compile_context, a, &_tmp_a, lhs_info, reinterpret_input_as_3d);
-
-    // Configure transpose kernel
-    _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info);
-
-    // Configure and tune matrix multiply kernel
-    _mm_kernel->configure(compile_context, &_tmp_a, &_tmp_b, c, output, alpha, beta, true, reshape_info, gemm_info.fp_mixed_precision(), gemm_info.activation_info());
-
-    CLScheduler::get().tune_kernel_static(*_mm_kernel);
-
-    // Request memory for LHS and RHS reshape matrix
-    _aux_mem[LhsReshape] = MemoryInfo(offset_int_vec(LhsReshape), MemoryLifetime::Temporary, _tmp_a.total_size());
-    _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size());
-}
-
-void ClGemm::configure_reshaped_v2(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta,
-                                   const GEMMInfo &gemm_info)
+void ClGemm::configure_reshaped(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta,
+                                const GEMMInfo &gemm_info)
 {
     DataType           data_type               = a->data_type();
     bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
@@ -311,7 +255,7 @@
 
     // Set the target for the kernels
     _reshape_lhs_kernel->set_target(gpu_target);
-    _mm_kernel->set_target(gpu_target);
+    _mm_reshaped_kernel->set_target(gpu_target);
 
     GEMMLHSMatrixInfo lhs_info{};
     GEMMRHSMatrixInfo rhs_info{};
@@ -354,7 +298,8 @@
     kernel_info.activation_info         = gemm_info.activation_info();
 
     // Set the target for the kernels
-    _mm_kernel->set_target(gpu_target);
+    _mm_reshaped_only_rhs_kernel->set_target(gpu_target);
+    _mm_reshaped_only_rhs_fallback_kernel->set_target(gpu_target);
 
     GEMMLHSMatrixInfo lhs_info{};
     GEMMRHSMatrixInfo rhs_info{};
@@ -381,78 +326,35 @@
     _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size());
 }
 
-Status ClGemm::validate_native_v1(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info)
+Status ClGemm::validate_native(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_UNUSED(alpha);
     ARM_COMPUTE_UNUSED(output);
 
     // Get the GPU target
     const GPUTarget    gpu_target              = CLScheduler::get().target();
+    DataType           data_type               = a->data_type();
     bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
     const unsigned int m                       = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
     const unsigned int n                       = b->dimension(0);
     const unsigned int k                       = a->dimension(0);
+    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();
+    const bool         broadcast_bias          = gemm_info.broadcast_bias();
 
-    const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d, gemm_info.broadcast_bias());
+    GEMMKernelInfo kernel_info;
+    kernel_info.m                       = m;
+    kernel_info.n                       = n;
+    kernel_info.k                       = k;
+    kernel_info.depth_output_gemm3d     = depth_output_gemm3d;
+    kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d;
+    kernel_info.broadcast_bias          = broadcast_bias;
+    kernel_info.activation_info         = gemm_info.activation_info();
+
+    auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size });
 
     // Validate matrix multiply
-    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyKernel::validate(a, b, c, output, alpha, beta,
-                                                                     false, reshape_info, gpu_target, gemm_info.fp_mixed_precision(), gemm_info.activation_info()));
-
-    return Status{};
-}
-
-Status ClGemm::validate_reshaped_v1(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info)
-{
-    ARM_COMPUTE_UNUSED(alpha);
-    ARM_COMPUTE_UNUSED(output);
-
-    TensorInfo tmp_a_info{};
-    TensorInfo tmp_b_info{};
-
-    // Get the GPU target
-    const GPUTarget    gpu_target                = CLScheduler::get().target();
-    const unsigned int m                         = gemm_info.reinterpret_input_as_3d() ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
-    const unsigned int n                         = b->dimension(0);
-    const unsigned int k                         = a->dimension(0);
-    int                mult_transpose1xW_width   = 1;
-    int                mult_interleave4x4_height = 1;
-    const int          depth_output_gemm3d       = gemm_info.depth_output_gemm3d();
-
-    if(get_arch_from_target(gpu_target) == GPUTarget::BIFROST)
-    {
-        mult_transpose1xW_width   = 4;
-        mult_interleave4x4_height = 2;
-    }
-
-    GEMMRHSMatrixInfo rhs_info;
-    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;
-
-    GEMMLHSMatrixInfo lhs_info;
-    lhs_info.m0         = 4;
-    lhs_info.k0         = 4;
-    lhs_info.v0         = mult_interleave4x4_height;
-    lhs_info.interleave = true;
-    lhs_info.transpose  = true;
-
-    const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d, false, gemm_info.broadcast_bias());
-
-    // 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()));
-
-    // Validate transpose kernel
-    auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
-    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info));
-
-    // Validate matrix multiply
-    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyKernel::validate(&tmp_a_info, &tmp_b_info, c, output, alpha, beta,
-                                                                     true, reshape_info, gpu_target, gemm_info.fp_mixed_precision(), gemm_info.activation_info()));
+    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyNativeKernel::validate(a, b, c, output, alpha, beta, config.lhs_info, config.rhs_info, kernel_info));
 
     return Status{};
 }
@@ -583,19 +485,14 @@
 
     switch(_gemm_kernel_type)
     {
-        case CLGEMMKernelType::NATIVE_V1:
+        case CLGEMMKernelType::NATIVE:
         {
-            configure_native_v1(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info);
-            break;
-        }
-        case CLGEMMKernelType::RESHAPED_V1:
-        {
-            configure_reshaped_v1(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info);
+            configure_native(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info);
             break;
         }
         case CLGEMMKernelType::RESHAPED:
         {
-            configure_reshaped_v2(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info);
+            configure_reshaped(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info);
             break;
         }
         case CLGEMMKernelType::RESHAPED_ONLY_RHS:
@@ -632,14 +529,9 @@
 
     switch(gemm_kernel_type)
     {
-        case CLGEMMKernelType::NATIVE_V1:
+        case CLGEMMKernelType::NATIVE:
         {
-            ARM_COMPUTE_RETURN_ON_ERROR(validate_native_v1(a, b, c_to_use, output, alpha, beta, gemm_info));
-            break;
-        }
-        case CLGEMMKernelType::RESHAPED_V1:
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped_v1(a, b, c_to_use, output, alpha, beta, gemm_info));
+            ARM_COMPUTE_RETURN_ON_ERROR(validate_native(a, b, c_to_use, output, alpha, beta, gemm_info));
             break;
         }
         case CLGEMMKernelType::RESHAPED:
@@ -679,12 +571,11 @@
     // Run matrix multiply kernel
     switch(_gemm_kernel_type)
     {
-        case CLGEMMKernelType::NATIVE_V1:
+        case CLGEMMKernelType::NATIVE:
         {
-            CLScheduler::get().enqueue_op(*_mm_kernel, tensors, true);
+            CLScheduler::get().enqueue_op(*_mm_native_kernel, tensors, true);
             break;
         }
-        case CLGEMMKernelType::RESHAPED_V1:
         case CLGEMMKernelType::RESHAPED:
         {
             // Run interleave kernel
@@ -704,10 +595,6 @@
             {
                 CLScheduler::get().enqueue_op(*_mm_reshaped_kernel, gemm_reshaped_pack, true);
             }
-            else
-            {
-                CLScheduler::get().enqueue_op(*_mm_kernel, gemm_reshaped_pack, true);
-            }
             break;
         }
         case CLGEMMKernelType::RESHAPED_ONLY_RHS:
diff --git a/src/gpu/cl/operators/ClGemm.h b/src/gpu/cl/operators/ClGemm.h
index 60bb78c..fd53648 100644
--- a/src/gpu/cl/operators/ClGemm.h
+++ b/src/gpu/cl/operators/ClGemm.h
@@ -31,7 +31,6 @@
 #include "src/gpu/cl/ClCompileContext.h"
 #include "src/gpu/cl/IClKernel.h"
 #include "src/gpu/cl/IClOperator.h"
-#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.h"
 #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h"
 #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h"
 #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h"
@@ -46,10 +45,10 @@
 {
 /** Basic function to execute GEMM on OpenCL. This function calls the following OpenCL kernels:
  *
- *  -# @ref kernels::ClGemmReshapeLhsMatrixKernel (only if the RESHAPED_V1 is selected by the heuristic model)
- *  -# @ref kernels::ClGemmReshapeRhsMatrixKernel (only if either the RESHAPED_V1 or RESHAPED_ONLY_RHS is selected by the select_gemm_kernel method())
- *  -# @ref kernels::ClGemmMatrixMultiplyKernel (only if either the NATIVE or RESHAPED_V1 is selected by the select_gemm_kernel method())
- *  -# @ref kernels::ClGemmMatrixMultiplyReshapedKernel (only if RESHAPED_V1 is selected by the select_gemm_kernel method())
+ *  -# @ref kernels::ClGemmReshapeLhsMatrixKernel (only if the RESHAPED is selected by the heuristic model)
+ *  -# @ref kernels::ClGemmReshapeRhsMatrixKernel (only if either the RESHAPED or RESHAPED_ONLY_RHS is selected by the select_gemm_kernel method())
+ *  -# @ref kernels::ClGemmMatrixMultiplyNativeKernel (only if NATIVE is selected by the select_gemm_kernel method())
+ *  -# @ref kernels::ClGemmMatrixMultiplyReshapedKernel (only if RESHAPED is selected by the select_gemm_kernel method())
  *  -# @ref kernels::ClGemmMatrixMultiplyReshapedOnlyRhsKernel (only if RESHAPED_ONLY_RHS is selected by the select_gemm_kernel method())
  */
 class ClGemm : public IClOperator
@@ -100,13 +99,11 @@
     experimental::MemoryRequirements workspace() const override;
 
 private:
-    void configure_native_v1(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
-    void configure_reshaped_v1(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
-    void configure_reshaped_v2(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
+    void configure_native(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
+    void configure_reshaped(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
     void configure_reshaped_only_rhs(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
 
-    static Status validate_native_v1(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
-    static Status validate_reshaped_v1(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
+    static Status validate_native(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
     static Status validate_reshaped(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
     static Status validate_reshaped_only_rhs(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
 
@@ -119,9 +116,9 @@
     };
 
 private:
-    std::unique_ptr<kernels::ClGemmMatrixMultiplyKernel>                _mm_kernel;
     std::unique_ptr<kernels::ClGemmReshapeLhsMatrixKernel>              _reshape_lhs_kernel;
     std::unique_ptr<kernels::ClGemmReshapeRhsMatrixKernel>              _reshape_rhs_kernel;
+    std::unique_ptr<kernels::ClGemmMatrixMultiplyNativeKernel>          _mm_native_kernel;
     std::unique_ptr<kernels::ClGemmMatrixMultiplyReshapedKernel>        _mm_reshaped_kernel;
     std::unique_ptr<kernels::ClGemmMatrixMultiplyReshapedOnlyRhsKernel> _mm_reshaped_only_rhs_kernel;
     std::unique_ptr<kernels::ClGemmMatrixMultiplyReshapedOnlyRhsKernel> _mm_reshaped_only_rhs_fallback_kernel;
diff --git a/src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.cpp b/src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.cpp
index 6fd7e52..7a62186 100644
--- a/src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.cpp
+++ b/src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.cpp
@@ -239,7 +239,6 @@
     GEMMLHSMatrixInfo lhs_info;
 
     // 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 unsigned int m                       = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
diff --git a/src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp b/src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp
index 67253c7..18ade97 100644
--- a/src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp
+++ b/src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp
@@ -125,13 +125,13 @@
 {
     ARM_COMPUTE_UNUSED(b);
 
-    CLGEMMKernelType gemm_type = CLGEMMKernelType::NATIVE_V1;
+    CLGEMMKernelType gemm_type = CLGEMMKernelType::NATIVE;
 
     if(is_rhs_constant)
     {
         if((m > 1) && (n < 16))
         {
-            gemm_type = CLGEMMKernelType::RESHAPED_V1;
+            gemm_type = CLGEMMKernelType::RESHAPED;
         }
         else if(m == 1)
         {
@@ -146,17 +146,17 @@
                 constexpr float fact1 = 1.66f;
                 constexpr float ops   = 12.0f;
                 const float     scale = k > 1024 ? 1.07f : 1.0f;
-                gemm_type             = (alpha + ((n * fact0) / ops) < ((fact1 * n * scale) / ops)) ? CLGEMMKernelType::RESHAPED_V1 : CLGEMMKernelType::NATIVE_V1;
+                gemm_type             = (alpha + ((n * fact0) / ops) < ((fact1 * n * scale) / ops)) ? CLGEMMKernelType::RESHAPED : CLGEMMKernelType::RESHAPED_ONLY_RHS;
             }
             else
             {
-                gemm_type = CLGEMMKernelType::NATIVE_V1;
+                gemm_type = CLGEMMKernelType::RESHAPED_ONLY_RHS;
             }
         }
 
         const auto workload = static_cast<float>((m * n) / 20.0f);
 
-        gemm_type = ((workload > 1600.0f) && (gemm_type == CLGEMMKernelType::RESHAPED_V1)) ? CLGEMMKernelType::RESHAPED : gemm_type;
+        gemm_type = ((workload > 1600.0f) && (gemm_type == CLGEMMKernelType::RESHAPED)) ? CLGEMMKernelType::RESHAPED : gemm_type;
     }
 
     return gemm_type;
@@ -179,7 +179,7 @@
     }
     else
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
 }
 
@@ -203,7 +203,7 @@
 
     if(!is_rhs_constant)
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
     if(m == 1)
     {
@@ -260,7 +260,7 @@
 
     if(!is_rhs_constant)
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
 
     if(m == 1)
@@ -387,7 +387,7 @@
 
     if(!is_rhs_constant)
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
 
     if(m == 1)
@@ -447,7 +447,7 @@
 {
     if(!is_rhs_constant)
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
 
     if(m == 1)
@@ -559,19 +559,14 @@
 CLGEMMKernelType CLGEMMDefaultTypeBifrost::g71_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
 {
     ARM_COMPUTE_UNUSED(b);
+    ARM_COMPUTE_UNUSED(n);
+    ARM_COMPUTE_UNUSED(k);
 
     if(is_rhs_constant)
     {
         if(m == 1)
         {
-            if(n > k)
-            {
-                return CLGEMMKernelType::NATIVE_V1;
-            }
-            else
-            {
-                return CLGEMMKernelType::RESHAPED_ONLY_RHS;
-            }
+            return CLGEMMKernelType::RESHAPED_ONLY_RHS;
         }
         else
         {
@@ -580,7 +575,7 @@
     }
     else
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
 }
 } // namespace cl_gemm
diff --git a/src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp b/src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp
index a64de99..ef30b28 100644
--- a/src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp
+++ b/src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp
@@ -73,7 +73,7 @@
     ARM_COMPUTE_UNUSED(n, k, b);
 
     // We reshape the matrices only if we do not have the vector-by-matrix case and we reshape the matrix B only once
-    return ((m != 1) && is_rhs_constant) ? CLGEMMKernelType::RESHAPED_V1 : CLGEMMKernelType::NATIVE_V1;
+    return ((m != 1) && is_rhs_constant) ? CLGEMMKernelType::RESHAPED : CLGEMMKernelType::NATIVE;
 }
 
 CLGEMMKernelType CLGEMMDefaultTypeMidgard::default_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
@@ -81,7 +81,7 @@
     ARM_COMPUTE_UNUSED(n, k, b);
 
     // We reshape the matrices only if we do not have the vector-by-matrix case and we reshape the matrix B only once
-    return ((m != 1) && is_rhs_constant) ? CLGEMMKernelType::RESHAPED_V1 : CLGEMMKernelType::NATIVE_V1;
+    return ((m != 1) && is_rhs_constant) ? CLGEMMKernelType::RESHAPED : CLGEMMKernelType::NATIVE;
 }
 
 CLGEMMKernelType CLGEMMDefaultTypeMidgard::default_q8(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
diff --git a/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp b/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
index b3403b2..64271a8 100644
--- a/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
+++ b/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
@@ -108,21 +108,21 @@
 {
     ARM_COMPUTE_UNUSED(m, n, k, b);
 
-    return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE_V1;
+    return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE;
 }
 
 CLGEMMKernelType CLGEMMDefaultTypeValhall::default_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
 {
     ARM_COMPUTE_UNUSED(m, n, k, b);
 
-    return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE_V1;
+    return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE;
 }
 
 CLGEMMKernelType CLGEMMDefaultTypeValhall::g77_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
 {
     if(!is_rhs_constant)
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
 
     if(m == 1)
@@ -242,7 +242,7 @@
 
     if(!is_rhs_constant)
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
 
     if(m == 1)
@@ -301,7 +301,7 @@
 
     if(!is_rhs_constant)
     {
-        return CLGEMMKernelType::NATIVE_V1;
+        return CLGEMMKernelType::NATIVE;
     }
 
     return CLGEMMKernelType::RESHAPED_ONLY_RHS;
diff --git a/tests/validate_examples/cl_gemm.cpp b/tests/validate_examples/cl_gemm.cpp
index 717ba77..82dfc05 100644
--- a/tests/validate_examples/cl_gemm.cpp
+++ b/tests/validate_examples/cl_gemm.cpp
@@ -39,7 +39,6 @@
 #include "src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
 #include "src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h"
 #include "src/core/CL/kernels/CLGEMMLowpReductionKernel.h"
-#include "src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h"
 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h"
 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
 #include "src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
diff --git a/tests/validation/CL/GEMMMatrixMultiply.cpp b/tests/validation/CL/GEMMMatrixMultiply.cpp
deleted file mode 100644
index faa2413..0000000
--- a/tests/validation/CL/GEMMMatrixMultiply.cpp
+++ /dev/null
@@ -1,339 +0,0 @@
-/*
- * Copyright (c) 2019-2021 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/KernelDescriptors.h"
-#include "arm_compute/core/Types.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "arm_compute/runtime/CL/CLTensor.h"
-#include "arm_compute/runtime/CL/CLTensorAllocator.h"
-#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.h"
-#include "tests/CL/CLAccessor.h"
-#include "tests/CL/Helper.h"
-#include "tests/PaddingCalculator.h"
-#include "tests/datasets/ShapeDatasets.h"
-#include "tests/framework/Asserts.h"
-#include "tests/framework/Macros.h"
-#include "tests/framework/datasets/Datasets.h"
-#include "tests/validation/Validation.h"
-#include "tests/validation/fixtures/GEMMFixture.h"
-
-namespace arm_compute
-{
-namespace test
-{
-namespace validation
-{
-using namespace arm_compute::misc::shape_calculator;
-using namespace arm_compute::opencl::kernels;
-
-// Create function for CLGEMMMatrixMultiplyKernel
-using CLGEMMMatrixMultiplyNative = CLSynthetizeOperator<ClGemmMatrixMultiplyKernel>;
-
-// Fixture for GEMMMatrixMultiplyValidationFixture
-template <typename T>
-using CLGEMMMatrixMultiplyNativeFixture = GEMMMatrixMultiplyValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>;
-
-// Fixture for GEMMMatrixMultiply3DValidationFixture
-template <typename T>
-using CLGEMMMatrixMultiplyNative3DFixture = GEMMMatrixMultiply3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>;
-
-namespace
-{
-// *INDENT-OFF*
-// clang-format off
-RelativeTolerance<float> rel_tolerance_f32(0.001f);
-constexpr float          abs_tolerance_f32(0.0001f);
-
-RelativeTolerance<half> rel_tolerance_f16(half(0.2));
-constexpr float         tolerance_num_f16 = 0.02f;
-
-/** Alpha values to test */
-const auto alpha_values = framework::dataset::make("alpha", {1.0f, -0.75f} );
-
-/** Beta values to test */
-const auto beta_values = framework::dataset::make("beta", {-0.35f, 0.0f} );
-
-/** M, N combinations to test
- *  1: Special 1x1 case
- *  2: Special multples of processor size in both dimensions
- *  3: Non multiples of processor size in both dimensions
- *  4: Special 1x1003 case
-*/
-const auto m_n_values = zip(
-    framework::dataset::make("M", {1, 16, 37, 1}),
-    framework::dataset::make("N", {1, 16, 51, 1003})
-    );
-
-/** N values to test */
-const auto n_values = framework::dataset::make("N", {51, 1003});
-
-/** K values to test */
-const auto k_values = framework::dataset::make("K", 23);
-
-/** 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);
-
-/** Batch size values to test */
-const auto b_values = framework::dataset::make("batch_size", 1, 3);
-
-/** Activation values to test */
-const auto act_values = framework::dataset::make("Activation",
-{
-    ActivationLayerInfo(),
-    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
-});
-
-/** Broadcast bias from vector to matrix */
-const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", { false, true } );
-
-/** GPU architectures values to test */
-const auto gpu_arch_values = framework::dataset::make("GPUArch",
-{
-    GPUTarget::MIDGARD,
-    GPUTarget::BIFROST
-});
-
-/** Data types values to test in the configuration */
-const auto data_type_values = framework::dataset::make("DataType",
-{
-    DataType::F32,
-    DataType::F16
-});
-
-/** M values to test */
-const auto fp16_mixed_precision_values = framework::dataset::make("fp16_mixed_precision", {true, false});
-} // namespace
-
-TEST_SUITE(CL)
-TEST_SUITE(GEMMMatrixMultiply)
-TEST_CASE(Negative, framework::DatasetMode::ALL)
-{
-    // Unsupported QASYMM8 data type
-    {
-        const auto lhs                       = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::QASYMM8);
-        const auto rhs                       = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::QASYMM8);
-        const auto out                       = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::QASYMM8);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = false;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Unsupported SIZE_T data type
-    {
-        const auto lhs                       = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::SIZET);
-        const auto rhs                       = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::SIZET);
-        const auto out                       = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::SIZET);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = false;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Mixed precision with F32
-    {
-        const auto lhs                       = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = false;
-        const GEMMReshapeInfo reshape_info  = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const bool fp_mixed_precision        = true;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Max number of dimensions LHS matrix
-    {
-        const auto lhs                       = TensorInfo(TensorShape(13U, 12U, 1U, 1U, 4U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = false;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Max number of dimensions RHS matrix
-    {
-        const auto lhs                       = TensorInfo(TensorShape(13U, 12U, 1U, 4U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(14U, 13U, 1U, 4U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(14U, 12U, 1U, 4U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = false;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Broadcast bias
-    {
-        const auto lhs                       = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::F16);
-        const auto rhs                       = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F16);
-        // The correct shape should be bias = TensorInfo(TensorShape(14U, 1U, 1U, 1U), 1, DataType::F32);
-        const auto bias                      = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F16);
-        const auto out                       = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F16);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = false;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, true);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const bool fp_mixed_precision        = false;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Invalid dimensions for the bias
-    {
-        const auto lhs                       = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F32);
-        // The correct shape should be bias = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32);
-        const auto bias                      = TensorInfo(TensorShape(14U, 8U, 1U, 1U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = false;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const bool fp_mixed_precision        = false;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Invalid dimensions for the output
-    {
-        const auto lhs                       = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F32);
-        // The correct shape should be out = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(14U, 7U, 1U, 1U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = false;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-}
-
-TEST_SUITE(Float)
-TEST_SUITE(FP32)
-
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(
-                                                                   m_n_values,
-                                                                   k_values),
-                                                                   b_values),
-                                                                   alpha_values),
-                                                                   beta_values),
-                                                                   broadcast_bias_values),
-                                                                   framework::dataset::make("fp16_mixed_precision", false)),
-                                                                   act_values),
-                                                                   framework::dataset::make("DataType", DataType::F32)),
-                                                                   gpu_arch_values))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
-}
-
-FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<float>, 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),
-                                                                   alpha_values),
-                                                                   beta_values),
-                                                                   broadcast_bias_values),
-                                                                   framework::dataset::make("fp16_mixed_precision", false)),
-                                                                   act_values),
-                                                                   framework::dataset::make("DataType", DataType::F32)),
-                                                                   gpu_arch_values))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
-}
-
-TEST_SUITE_END() // FP32
-
-TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(
-                                                                   m_n_values,
-                                                                   k_values),
-                                                                   b_values),
-                                                                   alpha_values),
-                                                                   beta_values),
-                                                                   broadcast_bias_values),
-                                                                   fp16_mixed_precision_values),
-                                                                   act_values),
-                                                                   framework::dataset::make("DataType", DataType::F16)),
-                                                                   gpu_arch_values))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
-}
-
-FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<half>, 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),
-                                                                   alpha_values),
-                                                                   beta_values),
-                                                                   broadcast_bias_values),
-                                                                   fp16_mixed_precision_values),
-                                                                   act_values),
-                                                                   framework::dataset::make("DataType", DataType::F16)),
-                                                                   gpu_arch_values))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
-}
-
-TEST_SUITE_END() // FP16
-TEST_SUITE_END() // Float
-TEST_SUITE_END() // GEMMMatrixMuliplty
-TEST_SUITE_END() // CL
-} // namespace validation
-} // namespace test
-} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/CL/GEMMMatrixMultiplyInterleavedTransposed.cpp b/tests/validation/CL/GEMMMatrixMultiplyInterleavedTransposed.cpp
deleted file mode 100644
index 9313ae3..0000000
--- a/tests/validation/CL/GEMMMatrixMultiplyInterleavedTransposed.cpp
+++ /dev/null
@@ -1,334 +0,0 @@
-/*
- * Copyright (c) 2019-2021 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/KernelDescriptors.h"
-#include "arm_compute/core/Types.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "arm_compute/runtime/CL/CLTensor.h"
-#include "arm_compute/runtime/CL/CLTensorAllocator.h"
-#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.h"
-#include "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h"
-#include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h"
-#include "tests/CL/CLAccessor.h"
-#include "tests/CL/Helper.h"
-#include "tests/PaddingCalculator.h"
-#include "tests/datasets/ShapeDatasets.h"
-#include "tests/framework/Asserts.h"
-#include "tests/framework/Macros.h"
-#include "tests/framework/datasets/Datasets.h"
-#include "tests/validation/Validation.h"
-#include "tests/validation/fixtures/GEMMFixture.h"
-
-namespace arm_compute
-{
-namespace test
-{
-namespace validation
-{
-using namespace arm_compute::misc::shape_calculator;
-using namespace arm_compute::opencl::kernels;
-
-// Create function for ClGemmReshapeLhsMatrixKernel
-using CLGEMMReshapeLHSMatrix = CLSynthetizeOperator<ClGemmReshapeLhsMatrixKernel>;
-
-// Create function for ClGemmReshapeRhsMatrixKernel
-using CLGEMMReshapeRHSMatrix = CLSynthetizeOperator<ClGemmReshapeRhsMatrixKernel>;
-
-// Create function for ClGemmMatrixMultiplyKernel
-using CLGEMMMatrixMultiplyReshaped = CLSynthetizeOperator<ClGemmMatrixMultiplyKernel>;
-
-// Fixture for GEMMMatrixMultiplyInterleavedTransposedValidationFixture
-template <typename T>
-using CLGEMMMatrixMultiplyReshapedFixture =
-    GEMMMatrixMultiplyInterleavedTransposedValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>;
-
-// Fixture for GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture
-template <typename T>
-using CLGEMMMatrixMultiplyReshaped3DFixture =
-    GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>;
-
-namespace
-{
-// *INDENT-OFF*
-// clang-format off
-RelativeTolerance<float> rel_tolerance_f32(0.001f);
-constexpr float          abs_tolerance_f32(0.0001f);
-
-RelativeTolerance<half> rel_tolerance_f16(half(0.2));
-constexpr float         tolerance_num_f16 = 0.02f;
-
-/** Alpha values to test */
-const auto alpha_values = framework::dataset::make("alpha", {1.0f, -0.75f} );
-
-/** Beta values to test */
-const auto beta_values = framework::dataset::make("beta", {-0.35f, 0.0f} );
-
-/** M, N combinations to test
- *  1: Special 1x1 case
- *  2: Special multples of processor size in both dimensions
- *  3: Non multiples of processor size in both dimensions
-*/
-const auto m_n_values = zip(
-    framework::dataset::make("M", {1, 16, 37}),
-    framework::dataset::make("N", {1, 16, 51})
-    );
-
-/** 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);
-
-/** 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);
-
-/** Batch size values to test */
-const auto b_values = framework::dataset::make("batch_size", 1, 3);
-
-/** Activation values to test */
-const auto act_values = framework::dataset::make("Activation",
-{
-    ActivationLayerInfo(),
-    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
-});
-
-/** V0 values to test */
-const auto v0_values = framework::dataset::make("V0", 2);
-
-/** H0 values to test */
-const auto h0_values = framework::dataset::make("H0", 4);
-
-/** Broadcast bias from vector to matrix */
-const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", {false, true} );
-
-/** GPU architectures values to test */
-const auto gpu_arch_values = framework::dataset::make("GPUArch",
-{
-    GPUTarget::MIDGARD,
-    GPUTarget::BIFROST
-});
-
-/** Data types values to test in the configuration */
-const auto data_type_values = framework::dataset::make("DataType",
-{
-    DataType::F32,
-    DataType::F16
-});
-
-/** M values to test */
-const auto fp16_mixed_precision_values = framework::dataset::make("fp16_mixed_precision", {true, false});
-} // namespace
-
-TEST_SUITE(CL)
-TEST_SUITE(GEMMMatrixMultiplyInterleavedTransposed)
-TEST_CASE(Negative, framework::DatasetMode::ALL)
-{
-    // The following tests are already integrated in the GEMMMatrixMultiply validation because
-    // in common with this validation
-    // - Unsupported QASYMM8 data type
-    // - Unsupported SIZE_T data type
-    // - Mixed precision with F32
-    // - Max number of dimensions LHS matrix
-    // - Max number of dimensions RHS matrix
-
-    // Invalid LHS dimensions
-    {
-        // The correct shape should be: lhs = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32);
-        const auto lhs                       = TensorInfo(TensorShape(256U, 2U, 1U, 1U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32);
-        const auto bias                      = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = true;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const bool fp_mixed_precision        = false;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Invalid RHS dimensions
-    {
-        const auto lhs                       = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32);
-        // The correct shape should be rhs = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(104U, 4U, 1U, 1U), 1, DataType::F32);
-        const auto bias                      = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = true;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const bool fp_mixed_precision        = false;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Broadcast bias
-    {
-        const auto lhs                       = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32);
-        // The correct shape should be bias = TensorInfo(TensorShape(24U, 1U, 1U, 1U), 1, DataType::F32);
-        const auto bias                      = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = true;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, true);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const bool fp_mixed_precision        = false;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Invalid dimensions for the bias
-    {
-        const auto lhs                       = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32);
-        // The correct shape should be bias = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        const auto bias                      = TensorInfo(TensorShape(25U, 16U, 1U, 1U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = true;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const bool fp_mixed_precision        = false;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-
-    // Invalid dimensions for the output
-    {
-        const auto lhs                       = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32);
-        const auto rhs                       = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32);
-        const auto bias                      = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        // The correct shape should be out = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32);
-        const auto out                       = TensorInfo(TensorShape(24U, 13U, 1U, 1U), 1, DataType::F32);
-        constexpr float alpha                = 1.3f;
-        constexpr float beta                 = 0.7f;
-        const bool is_interleaved_transposed = true;
-        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, false);
-        const GPUTarget gpu_target           = GPUTarget::MIDGARD;
-        const bool fp_mixed_precision        = false;
-        const auto status    = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision);
-        ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
-    }
-}
-
-TEST_SUITE(Float)
-TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<float>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
-                                                                   m_n_values,
-                                                                   k_values),
-                                                                   b_values),
-                                                                   alpha_values),
-                                                                   beta_values),
-                                                                   v0_values),
-                                                                   h0_values),
-                                                                   broadcast_bias_values),
-                                                                   framework::dataset::make("fp16_mixed_precision", false)),
-                                                                   act_values),
-                                                                   framework::dataset::make("DataType", DataType::F32)),
-                                                                   gpu_arch_values))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
-}
-
-FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
-                                                                   m_w_values,
-                                                                   m_h_values),
-                                                                   n_values),
-                                                                   k_values),
-                                                                   b_values),
-                                                                   alpha_values),
-                                                                   beta_values),
-                                                                   v0_values),
-                                                                   h0_values),
-                                                                   broadcast_bias_values),
-                                                                   framework::dataset::make("fp16_mixed_precision", false)),
-                                                                   act_values),
-                                                                   framework::dataset::make("DataType", DataType::F32)),
-                                                                   gpu_arch_values))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
-}
-
-TEST_SUITE_END() // FP32
-
-TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<half>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
-                                                                   m_n_values,
-                                                                   k_values),
-                                                                   b_values),
-                                                                   alpha_values),
-                                                                   beta_values),
-                                                                   v0_values),
-                                                                   h0_values),
-                                                                   broadcast_bias_values),
-                                                                   fp16_mixed_precision_values),
-                                                                   act_values),
-                                                                   framework::dataset::make("DataType", DataType::F16)),
-                                                                   gpu_arch_values))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
-}
-
-FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
-                                                                   m_w_values,
-                                                                   m_h_values),
-                                                                   n_values),
-                                                                   k_values),
-                                                                   b_values),
-                                                                   alpha_values),
-                                                                   beta_values),
-                                                                   v0_values),
-                                                                   h0_values),
-                                                                   broadcast_bias_values),
-                                                                   fp16_mixed_precision_values),
-                                                                   act_values),
-                                                                   framework::dataset::make("DataType", DataType::F16)),
-                                                                   gpu_arch_values))
-{
-    // Validate output
-    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
-}
-
-TEST_SUITE_END() // FP16
-TEST_SUITE_END() // Float
-TEST_SUITE_END() // GEMMMatrixMulipltyInterleavedTransposed
-TEST_SUITE_END() // CL
-} // namespace validation
-} // namespace test
-} // namespace arm_compute
\ No newline at end of file
diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h
index 3e73b90..220c3ac 100644
--- a/utils/TypePrinter.h
+++ b/utils/TypePrinter.h
@@ -2360,14 +2360,6 @@
 {
     switch(val)
     {
-        case CLGEMMKernelType::NATIVE_V1:
-        {
-            return "Native_V1";
-        }
-        case CLGEMMKernelType::RESHAPED_V1:
-        {
-            return "Reshaped_V1";
-        }
         case CLGEMMKernelType::NATIVE:
         {
             return "Native";