COMPMID-1774: Implement CLGEMMReshapeLHSMatrixKernel to reshape the LHS matrix of GEMM/GEMMLowp

Change-Id: I8c5fd4c8bcdffda1522c83158981ed92baa045f4
Reviewed-on: https://review.mlplatform.org/364
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 466 insertions, 2 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index da85472005..7b98e5ae80 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -282,6 +282,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemm_mm_floating_point_f32_bifrost_1000", "gemm.cl" },
     { "gemm_lc_vm_f32", "gemm.cl" },
     { "gemm_transpose1xW", "gemm.cl" },
+    { "gemm_reshape_lhs_matrix_nt", "gemm.cl" },
     { "gemmlowp_matrix_a_reduction", "gemmlowp.cl" },
     { "gemmlowp_matrix_a_reduction_dot8", "gemmlowp.cl" },
     { "gemmlowp_matrix_b_reduction", "gemmlowp.cl" },
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index 7de15d018a..cf1e021929 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -23,6 +23,235 @@
  */
 #include "helpers.h"
 
+#if defined(M0) && defined(K0) && defined(V0) && defined(DATA_TYPE)
+
+/** This OpenCL kernel reshapes the lhs input matrix. The kernel splits the input matrix in blocks of size M0xK0 and stores each one (not transposed) in
+ *  the output matrix unrolling the values.
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=float)
+ * @note The block's dimensions (M0 and K0) must be passed at compile time using -DM0 and -DK0 (i.e. -DM0=2, -DK0=2).
+ * @note The number of M0xK0 vertical blocks to store on the same output row must be passed at compile time using -DV0 (i.e. -DV0=2)
+ * @note Only the following values for M0, K0 and V0 are supported:
+ *                                      M0: 2,3,4,5,6,7,8
+ *                                      K0: 2,4,8,16
+ *                                      V0: greater than 0
+ * @note In case the input has to be reinterpreted as a 3D tensor (i.e. input of convolution layer 1x1), the following information must be passed at compile time:
+ *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
+ *       -# HEIGHT_GEMM3D: The height of the input in case it has to be reinterpreted as a 3D tensor.
+ *       -# DEPTH_GEMM3D: The depth of the input in case it has to be reinterpreted as a 3D tensor
+ *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
+ * @note If the M0xK0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time.
+ *
+ * @param[in]  src_ptr                           Pointer to the source LHS tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  src_stride_x                      Stride of the source LHS tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source LHS tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                      Stride of the source LHS tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source LHS tensor
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
+ * @param[in]  cross_plane_pad                   (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
+ */
+__kernel void gemm_reshape_lhs_matrix_nt(TENSOR3D_DECLARATION(src),
+                                         TENSOR3D_DECLARATION(dst)
+#if defined(REINTERPRET_INPUT_AS_3D)
+                                         ,
+                                         uint cross_plane_pad
+#endif // REINTERPRET_INPUT_AS_3D
+                                        )
+{
+// Block size
+#define BLOCK_SIZE ((M0) * (K0))
+
+// Output offset X
+#if defined(INTERLEAVE)
+#define OUTPUT_OFFSET_X (K0)
+#else // defined(INTERLEAVE)
+#define OUTPUT_OFFSET_X (BLOCK_SIZE)
+#endif // defined(INTERLEAVE)
+
+// Output step X
+#if defined(INTERLEAVE)
+#define OUTPUT_STEP_X (K0) * (V0)
+#else // Do not interleave
+#define OUTPUT_STEP_X (K0)
+#endif // defined(INTERLEAVE)
+
+    // Compute source and destination addresses
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    // ------------------ Compute input/output addresses ---------------------------
+
+    // Compute the input address
+    __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)K0 * sizeof(DATA_TYPE) + y * (uint)M0 * src_stride_y;
+
+    // Compute the output address
+    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)BLOCK_SIZE * (uint)V0 * sizeof(DATA_TYPE)) + ((y / (uint)V0) * (uint)dst_stride_y) + ((y % V0) *
+                                 (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE));
+
+    uint zin0 = 0;
+    uint zin1 = 0;
+    uint zin2 = 0;
+    uint zin3 = 0;
+    uint zin4 = 0;
+    uint zin5 = 0;
+    uint zin6 = 0;
+    uint zin7 = 0;
+
+#if defined(REINTERPRET_INPUT_AS_3D)
+    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
+    // multiply src_stride_z by DEPTH_GEMM3D
+
+    // Note for the REINTERPRET_INPUT_AS_3D case
+    // 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      |
+    //  |                  |
+    //  |__________________|
+
+    input_ptr += z * (uint)src_stride_z * DEPTH_GEMM3D;
+
+    // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    zin0 = (0 + (uint)(y * M0)) / (uint)HEIGHT_GEMM3D;
+    zin0 = min((uint)(DEPTH_GEMM3D - 1), zin0);
+    zin0 *= (cross_plane_pad * src_stride_y);
+#if M0 > 1
+    zin1 = (1 + (uint)(y * M0)) / (uint)HEIGHT_GEMM3D;
+    zin1 = min((uint)(DEPTH_GEMM3D - 1), zin1);
+    zin1 *= (cross_plane_pad * src_stride_y);
+#endif // M0 > 1
+#if M0 > 2
+    zin2 = (2 + (uint)(y * M0)) / (uint)HEIGHT_GEMM3D;
+    zin2 = min((uint)(DEPTH_GEMM3D - 1), zin2);
+    zin2 *= (cross_plane_pad * src_stride_y);
+#endif // M0 > 2
+#if M0 > 3
+    zin3 = (3 + (uint)(y * M0)) / (uint)HEIGHT_GEMM3D;
+    zin3 = min((uint)(DEPTH_GEMM3D - 1), zin3);
+    zin3 *= (cross_plane_pad * src_stride_y);
+#endif // M0 > 3
+#if M0 > 4
+    zin4 = (4 + (uint)(y * M0)) / (uint)HEIGHT_GEMM3D;
+    zin4 = min((uint)(DEPTH_GEMM3D - 1), zin4);
+    zin4 *= (cross_plane_pad * src_stride_y);
+#endif // M0 > 4
+#if M0 > 5
+    zin5 = (5 + (uint)(y * M0)) / (uint)HEIGHT_GEMM3D;
+    zin5 = min((uint)(DEPTH_GEMM3D - 1), zin5);
+    zin5 *= (cross_plane_pad * src_stride_y);
+#endif // M0 > 5
+#if M0 > 6
+    zin6 = (6 + (uint)(y * M0)) / (uint)HEIGHT_GEMM3D;
+    zin6 = min((uint)(DEPTH_GEMM3D - 1), zin6);
+    zin6 *= (cross_plane_pad * src_stride_y);
+#endif // M0 > 6
+#if M0 > 6
+    zin7 = (7 + (uint)(y * M0)) / (uint)HEIGHT_GEMM3D;
+    zin7 = min((uint)(DEPTH_GEMM3D - 1), zin7);
+    zin7 *= (cross_plane_pad * src_stride_y);
+#endif // M0 > 7
+
+#else // defined(REINTERPRET_INPUT_AS_3D)
+
+    input_ptr += z * (uint)src_stride_z;
+
+#endif // defined(REINTERPRET_INPUT_AS_3D)
+
+    // Add offset for batched GEMM
+    output_ptr += z * (uint)dst_stride_z;
+
+    // ---------------------------Load input values --------------------------------
+
+    // Load values from the LHS matrix
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    a0 = VLOAD(K0)(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y + zin0));
+#if M0 > 1
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    a1 = VLOAD(K0)(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y + zin1));
+#endif // M0 > 1
+#if M0 > 2
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    a2 = VLOAD(K0)(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y + zin2));
+#endif // M0 > 2
+#if M0 > 3
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    a3 = VLOAD(K0)(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y + zin3));
+#endif // M0 > 3
+#if M0 > 4
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    a4 = VLOAD(K0)(0, (__global DATA_TYPE *)(input_ptr + 4 * src_stride_y + zin4));
+#endif // M0 > 4
+#if M0 > 5
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    a5 = VLOAD(K0)(0, (__global DATA_TYPE *)(input_ptr + 5 * src_stride_y + zin5));
+#endif // M0 > 5
+#if M0 > 6
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    a6 = VLOAD(K0)(0, (__global DATA_TYPE *)(input_ptr + 6 * src_stride_y + zin6));
+#endif // M0 > 6
+#if M0 > 7
+    VEC_DATA_TYPE(DATA_TYPE, K0)
+    a7 = VLOAD(K0)(0, (__global DATA_TYPE *)(input_ptr + 7 * src_stride_y + zin7));
+#endif // M0 > 7
+
+    // ---------------------------Store output values ------------------------------
+
+    VSTORE(K0)
+    (a0, 0, (__global DATA_TYPE *)(output_ptr + 0 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#if M0 > 1
+    VSTORE(K0)
+    (a1, 0, (__global DATA_TYPE *)(output_ptr + 1 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 1
+#if M0 > 2
+    VSTORE(K0)
+    (a2, 0, (__global DATA_TYPE *)(output_ptr + 2 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 2
+#if M0 > 3
+    VSTORE(K0)
+    (a3, 0, (__global DATA_TYPE *)(output_ptr + 3 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 3
+#if M0 > 4
+    VSTORE(K0)
+    (a4, 0, (__global DATA_TYPE *)(output_ptr + 4 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 4
+#if M0 > 5
+    VSTORE(K0)
+    (a5, 0, (__global DATA_TYPE *)(output_ptr + 5 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 5
+#if M0 > 6
+    VSTORE(K0)
+    (a6, 0, (__global DATA_TYPE *)(output_ptr + 6 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 6
+#if M0 > 7
+    VSTORE(K0)
+    (a7, 0, (__global DATA_TYPE *)(output_ptr + 7 * OUTPUT_STEP_X * sizeof(DATA_TYPE)));
+#endif // M0 > 7
+
+#undef BLOCK_SIZE
+#undef OUTPUT_OFFSET_X
+#undef OUTPUT_STEP_X
+}
+#endif // defined(M0) && defined(K0) && defined(V0) && defined(DATA_TYPE)
+
 #if defined(TRANSPOSE_W) && defined(MULT_TRANSPOSE1XW_WIDTH)
 
 #if ELEMENT_SIZE == 1
@@ -193,7 +422,7 @@ __kernel void gemm_interleave4x4(TENSOR3D_DECLARATION(src),
     vstore4(a1, 0, ((__global DATA_TYPE *)(dst_ptr + dst_addr_in_bytes) + 4 * MULT_INTERLEAVE4X4_HEIGHT));
     vstore4(a2, 0, ((__global DATA_TYPE *)(dst_ptr + dst_addr_in_bytes) + 8 * MULT_INTERLEAVE4X4_HEIGHT));
     vstore4(a3, 0, ((__global DATA_TYPE *)(dst_ptr + dst_addr_in_bytes) + 12 * MULT_INTERLEAVE4X4_HEIGHT));
-#else // defined(UNROLL_BLOCK)
+#else  // defined(UNROLL_BLOCK)
     VEC_DATA_TYPE(DATA_TYPE, 4)
     val0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s0, a1.s0, a2.s0, a3.s0);
     vstore4(val0, 0, ((__global DATA_TYPE *)(dst_ptr + dst_addr_in_bytes) + 0 * MULT_INTERLEAVE4X4_HEIGHT));
diff --git a/src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.cpp b/src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.cpp
new file mode 100644
index 0000000000..e0af5801a8
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.cpp
@@ -0,0 +1,221 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, bool reinterpret_input_as_3d)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.transpose);
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.v0 == 0);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((lhs_info.k0 & (lhs_info.k0 - 1)), "Only power of two values are allowed for k0");
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::U8, DataType::S8,
+                                                         DataType::U16, DataType::S16, DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_lhs_reshaped_shape(*input, lhs_info, reinterpret_input_as_3d));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, bool reinterpret_input_as_3d)
+{
+    const unsigned int num_elems_processed_per_iteration_x = lhs_info.k0;
+    const unsigned int num_elems_processed_per_iteration_y = lhs_info.m0;
+    bool               window_changed                      = false;
+
+    TensorInfo tmp_info(*input);
+
+    if(reinterpret_input_as_3d)
+    {
+        // Since the input tensor has to be reinterpreted as 3D and the execute window is based on a 2D interleave,
+        // the window needs to be constructed on the 2D collapsed version of the tensor
+        TensorShape tmp_shape(input->tensor_shape());
+        tmp_shape.collapse(2U, 1U);
+        tmp_info.set_tensor_shape(tmp_shape);
+    }
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*input, lhs_info, reinterpret_input_as_3d)));
+
+    // Configure window
+    // Note: bottom paddings are calculated manually as the input can be reinterpreted as 3D tensor
+    // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic
+    const int m          = reinterpret_input_as_3d ? input->tensor_shape()[1] * input->tensor_shape()[2] : input->tensor_shape()[1];
+    const int bottom_pad = ceil_to_multiple(m, num_elems_processed_per_iteration_y) - m;
+
+    Window win    = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+    Window win_in = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+    AccessWindowStatic input_access(input, 0, 0,
+                                    ceil_to_multiple(input->dimension(0), num_elems_processed_per_iteration_x),
+                                    input->dimension(1) + bottom_pad);
+    AccessWindowStatic output_access(output, 0, 0, output->dimension(0), output->dimension(1));
+
+    window_changed = update_window_and_padding(win_in, input_access) || // window used by the execute_window_loop
+                     update_window_and_padding(win, output_access);     // window used to update the padding requirements of output tensor
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
+
+    // Collapse along the Z direction
+    // This collapse needs to be here in order to tune the Z dimension of LWS
+    Window collapsed = win.collapse(win, Window::DimZ);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, collapsed);
+}
+} // namespace
+
+CLGEMMReshapeLHSMatrixKernel::CLGEMMReshapeLHSMatrixKernel()
+    : _input(nullptr), _output(nullptr), _reinterpret_input_as_3d(false)
+{
+}
+
+void CLGEMMReshapeLHSMatrixKernel::configure(const ICLTensor *input, ICLTensor *output, const GEMMLHSMatrixInfo &lhs_info, bool reinterpret_input_as_3d)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), lhs_info, reinterpret_input_as_3d));
+
+    _input                   = input;
+    _output                  = output;
+    _reinterpret_input_as_3d = reinterpret_input_as_3d;
+
+    // Create build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
+    build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));
+    build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));
+    build_opts.add_option_if(lhs_info.interleave, "-DINTERLEAVE");
+    build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
+    build_opts.add_option_if(_reinterpret_input_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(input->info()->dimension(1)));
+    build_opts.add_option_if(_reinterpret_input_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(input->info()->dimension(2)));
+
+    switch(input->info()->element_size())
+    {
+        case 1:
+            build_opts.add_option("-DDATA_TYPE=uchar");
+            break;
+        case 2:
+            build_opts.add_option("-DDATA_TYPE=ushort");
+            break;
+        case 4:
+            build_opts.add_option("-DDATA_TYPE=uint");
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Data type not supported");
+    }
+
+    std::string kernel_name("gemm_reshape_lhs_matrix_");
+    kernel_name += lhs_info.transpose ? "t" : "nt";
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info(), lhs_info, reinterpret_input_as_3d);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
+    // Set config_id for enabling LWS tuning
+    _config_id = "gemm_reshape_lhs_matrix_";
+    _config_id += (_reinterpret_input_as_3d ? "3d_" : "");
+    _config_id += lower_string(string_from_data_type(input->info()->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(2));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(lhs_info.m0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(lhs_info.k0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(lhs_info.v0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(lhs_info.interleave);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(lhs_info.transpose);
+}
+
+Status CLGEMMReshapeLHSMatrixKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, bool reinterpret_input_as_3d)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, lhs_info, reinterpret_input_as_3d));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), lhs_info, reinterpret_input_as_3d).first);
+
+    return Status{};
+}
+
+void CLGEMMReshapeLHSMatrixKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window slice = window.first_slice_window_3D();
+
+    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                  = 2 * num_arguments_per_3D_tensor();
+        const unsigned int total_cross_plane_pad = _input->info()->padding().top + _input->info()->padding().bottom;
+        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
+    }
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice);
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(window.slide_window_slice_3D(slice));
+}
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLReverseKernel.cpp b/src/core/CL/kernels/CLReverseKernel.cpp
index 2859a51ce1..adbdb11c5f 100644
--- a/src/core/CL/kernels/CLReverseKernel.cpp
+++ b/src/core/CL/kernels/CLReverseKernel.cpp
@@ -80,7 +80,20 @@ void CLReverseKernel::configure(const ICLTensor *input, ICLTensor *output, const
     // Set kernel build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DNUM_REVERSE_DIMS=" + support::cpp11::to_string(axis->info()->dimension(0)));
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    switch(input->info()->element_size())
+    {
+        case 1:
+            build_opts.add_option("-DDATA_TYPE=uchar");
+            break;
+        case 2:
+            build_opts.add_option("-DDATA_TYPE=ushort");
+            break;
+        case 4:
+            build_opts.add_option("-DDATA_TYPE=uint");
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Data type not supported");
+    }
 
     // Create kernel
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("reverse", build_opts.options()));