Port CLGEMM to memory injecting interface

Moves the following kernels:
 - CLGEMMMatrixMultiplyKernel
 - CLGEMMMatrixMultiplyNativeKernel
 - CLGEMMMatrixMultipluReshapedKernel
 - CLGEMMMatrixMultiplyReshapedOnlyRHSKernel

 Moves the following functions
 - CLGEMM

Introduces facilities to easy handling of auxiliary temporary buffers
under then new run interface. Such are:
 - CLAuxTensorHandler: That allows wrapping of workspace buffers memory
 to CLBuffer objects
 - Ability to inject TensorInfo to allocator without transferring
 ownership. This reduce the copy overhead if needed.

Resolves: COMPMID-4188

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I7055435d831b05b749b26302082e4ac45f26dfb0
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5498
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

20 files changed, 1084 insertions, 894 deletions

diff --git a/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp
index 00d9a9ec89..8d1a91e420 100644
--- a/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp
@@ -31,7 +31,6 @@
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
 #include "src/core/CL/kernels/CLFillBorderKernel.h"
-#include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 #include "src/core/CL/kernels/CLWeightsReshapeKernel.h"
 #include "src/core/helpers/AutoConfiguration.h"
 
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 945675f4dd..991472bb7a 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -35,11 +35,6 @@
 #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"
-#include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 #include "src/core/gpu/cl/kernels/ClTransposeKernel.h"
 #include "support/Cast.h"
 
diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp
index cf1a82bc5a..1bc785a0a7 100644
--- a/src/runtime/CL/functions/CLGEMM.cpp
+++ b/src/runtime/CL/functions/CLGEMM.cpp
@@ -23,646 +23,48 @@
  */
 #include "arm_compute/runtime/CL/functions/CLGEMM.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/Error.h"
-#include "arm_compute/core/GPUTarget.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/KernelDescriptors.h"
-#include "arm_compute/core/Log.h"
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Utils.h"
-#include "arm_compute/core/Validate.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "arm_compute/runtime/CL/CLScheduler.h"
-#include "arm_compute/runtime/ITensorAllocator.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"
-#include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
-#include "src/core/helpers/AutoConfiguration.h"
-#include "src/core/utils/helpers/float_ops.h"
-#include "src/runtime/CL/gemm/CLGEMMKernelSelection.h"
-#include "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.h"
-#include "support/Cast.h"
-#include "utils/TypePrinter.h"
+#include "arm_compute/runtime/CL/functions/CLGEMM.h"
+#include "src/core/helpers/MemoryHelpers.h"
+#include "src/runtime/gpu/cl/operators/ClGemm.h"
 
 namespace arm_compute
 {
-using namespace arm_compute::misc::shape_calculator;
-using namespace arm_compute::cl_gemm;
-using namespace arm_compute::utils::cast;
-
-namespace weights_transformations
-{
-CLGEMMReshapeRHSMatrixKernelManaged::CLGEMMReshapeRHSMatrixKernelManaged()
-    : _kernel(std::make_unique<CLGEMMReshapeRHSMatrixKernel>())
-{
-}
-
-CLGEMMReshapeRHSMatrixKernelManaged::~CLGEMMReshapeRHSMatrixKernelManaged() = default;
-
-void CLGEMMReshapeRHSMatrixKernelManaged::run()
-{
-    _output.allocator()->allocate();
-    CLScheduler::get().enqueue(*_kernel, false);
-    _reshape_run = true;
-}
-
-void CLGEMMReshapeRHSMatrixKernelManaged::release()
-{
-    _output.allocator()->free();
-}
-
-ICLTensor *CLGEMMReshapeRHSMatrixKernelManaged::get_weights()
-{
-    return &_output;
-}
-
-uint32_t CLGEMMReshapeRHSMatrixKernelManaged::uid()
-{
-    return _uid;
-}
-
-void CLGEMMReshapeRHSMatrixKernelManaged::configure(const ICLTensor *input, GEMMRHSMatrixInfo info)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), input, info);
-}
-
-void CLGEMMReshapeRHSMatrixKernelManaged::configure(const CLCompileContext &compile_context, const ICLTensor *input, GEMMRHSMatrixInfo info)
-{
-    _kernel->configure(compile_context, input, &_output, info);
-}
-} // namespace weights_transformations
-
-namespace
-{
-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;
-        }
-    }
-}
-//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)
-{
-    auto gemm_kernel = auto_heuristics::select_mlgo_gemm_kernel(query, reshape_b_only_on_first_run);
-    if(bool(gemm_kernel))
-    {
-        if(validate_gemm_kernel(gemm_kernel.gemm_type))
-        {
-            ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from mlgo heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str());
-            return gemm_kernel.gemm_type;
-        }
-    }
-    gemm_kernel = auto_heuristics::select_default_gemm_kernel(query, reshape_b_only_on_first_run);
-    ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from default heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str());
-    return gemm_kernel.gemm_type;
-}
-// Validate lhs_info and rhs_info for reshaped only rhs kernel
-inline bool validate_lhs_rhs_info_reshaped_only_rhs(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c,
-                                                    const ITensorInfo *output, GEMMKernelInfo gemm_kernel_info)
-{
-    // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel
-    TensorInfo tmp_b_info{};
-    // Validate reshape RHS kernel
-    auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
-    if(!bool(CLGEMMReshapeRHSMatrixKernel::validate(b, &tmp_b_info, rhs_info)))
-    {
-        return false;
-    }
-    // Validate mm kernel
-    gemm_kernel_info.lhs_info  = lhs_info;
-    gemm_kernel_info.rhs_info  = rhs_info;
-    gemm_kernel_info.has_pad_y = false;
-    if(!bool(CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::validate(a, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info)))
-    {
-        return false;
-    }
-    gemm_kernel_info.has_pad_y = true;
-    if(!bool(CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::validate(a, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info)))
-    {
-        return false;
-    }
-    return true;
-}
-
-//Automatically select between mlgo (prioritized) and default heuristics for reshaped only rhs kernel configs
-inline std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery query, GEMMKernelInfo kernel_info, const ITensorInfo *a,
-                                                                                                 const ITensorInfo *b,
-                                                                                                 const ITensorInfo *c, const ITensorInfo *output)
-{
-    auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(query);
-    if(config)
-    {
-        if(validate_lhs_rhs_info_reshaped_only_rhs(config.lhs_info, config.rhs_info, a, b, c, output, kernel_info))
-        {
-            ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
-            return { config.lhs_info, config.rhs_info };
-        }
-    }
-    config = auto_heuristics::select_default_gemm_config_reshaped_only_rhs(query);
-    ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
-    return { config.lhs_info, config.rhs_info };
-}
-
-// Validate lhs_info and rhs_info for reshaped kernel
-inline bool validate_lhs_rhs_info_reshaped(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c,
-                                           const ITensorInfo *output, GEMMKernelInfo gemm_kernel_info, bool reinterpret_input_as_3d)
-{
-    // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped kernel
-    TensorInfo tmp_a_info{};
-    TensorInfo tmp_b_info{};
-
-    // Validate reshape LHS kernel
-    auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, reinterpret_input_as_3d)));
-    if(!bool(CLGEMMReshapeLHSMatrixKernel::validate(a, &tmp_a_info, lhs_info, reinterpret_input_as_3d)))
-    {
-        return false;
-    }
-
-    // Validate reshape RHS kernel
-    auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
-    if(!bool(CLGEMMReshapeRHSMatrixKernel::validate(b, &tmp_b_info, rhs_info)))
-    {
-        return false;
-    }
-    // Validate mm kernel
-    gemm_kernel_info.lhs_info = lhs_info;
-    gemm_kernel_info.rhs_info = rhs_info;
-    if(!bool(CLGEMMMatrixMultiplyReshapedKernel::validate(&tmp_a_info, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info)))
-    {
-        return false;
-    }
-    return true;
-}
-
-//Automatically select between mlgo (prioritized) and default heuristics for reshaped kernel configs
-inline std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped(auto_heuristics::CommonQuery query, GEMMKernelInfo kernel_info, const ITensorInfo *a, const ITensorInfo *b,
-                                                                                        const ITensorInfo *c, const ITensorInfo *output, bool reinterpret_input_as_3d)
-{
-    auto config = auto_heuristics::select_mlgo_gemm_config_reshaped(query);
-    if(config)
-    {
-        if(validate_lhs_rhs_info_reshaped(config.lhs_info, config.rhs_info, a, b, c, output, kernel_info, reinterpret_input_as_3d))
-        {
-            ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
-            return { config.lhs_info, config.rhs_info };
-        }
-    }
-    config = auto_heuristics::select_default_gemm_config_reshaped(query);
-    ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
-    return { config.lhs_info, config.rhs_info };
-}
-
-} // namespace
+using namespace arm_compute::experimental;
+using OperatorType = opencl::ClGemm;
+
+struct CLGEMM::Impl
+{
+    const ICLTensor              *a{ nullptr };
+    const ICLTensor              *b{ nullptr };
+    const ICLTensor              *c{ nullptr };
+    ICLTensor                    *dst{ nullptr };
+    std::unique_ptr<OperatorType> op{ nullptr };
+    MemoryGroup                   memory_group{};
+    IWeightsManager              *weights_manager{ nullptr };
+    CLTensor                      weights_transformed{};
+    ITensorPack                   run_pack{};
+    ITensorPack                   prep_pack{};
+    MemoryRequirements            aux_mem_req{};
+    WorkspaceData<CLTensor>       workspace_tensors{};
+    bool                          _is_prepared{ false };
+};
 
 CLGEMM::CLGEMM(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
-    : _memory_group(std::move(memory_manager)),
-      _weights_manager(weights_manager),
-      _mm_kernel(std::make_unique<CLGEMMMatrixMultiplyKernel>()),
-      _reshape_lhs_kernel(std::make_unique<CLGEMMReshapeLHSMatrixKernel>()),
-      _reshape_rhs_kernel(std::make_unique<CLGEMMReshapeRHSMatrixKernel>()),
-      _reshape_rhs_kernel_managed(std::make_unique<weights_transformations::CLGEMMReshapeRHSMatrixKernelManaged>()),
-      _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(),
-      _original_b(nullptr),
-      _lhs(nullptr),
-      _dst(nullptr),
-      _reshape_b_only_on_first_run(false),
-      _is_prepared(false),
-      _gemm_kernel_type(CLGEMMKernelType::NATIVE_V1)
+    : _impl(std::make_unique<Impl>())
 {
+    _impl->memory_group    = MemoryGroup(memory_manager);
+    _impl->weights_manager = weights_manager;
 }
 
 CLGEMM::~CLGEMM() = default;
 
-void CLGEMM::configure_native_v1(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta,
-                                 const GEMMInfo &gemm_info)
-{
-    const unsigned int m          = gemm_info.reinterpret_input_as_3d() ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
-    const unsigned int n          = b->info()->dimension(0);
-    const unsigned int k          = a->info()->dimension(0);
-    const GPUTarget    gpu_target = CLScheduler::get().target();
-
-    // Set the target for the kernels
-    _mm_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());
-
-    // 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);
-}
-
-void CLGEMM::configure_reshaped_v1(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *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->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
-    const unsigned int n                         = b->info()->dimension(0);
-    const unsigned int k                         = a->info()->dimension(0);
-    const int          depth_output_gemm3d       = gemm_info.depth_output_gemm3d();
-    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->info()->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());
-
-    const bool use_mm_b = (!_weights_manager || !_weights_manager->are_weights_managed(b));
-
-    // Manage intermediate buffers
-    _memory_group.manage(&_tmp_a);
-
-    if(!_reshape_b_only_on_first_run && use_mm_b)
-    {
-        _memory_group.manage(&_tmp_b);
-    }
-
-    // Configure interleave kernel
-    _reshape_lhs_kernel->configure(compile_context, a, &_tmp_a, lhs_info, reinterpret_input_as_3d);
-
-    // Configure transpose kernel
-    ICLTensor *reshaped_rhs = &_tmp_b;
-    if(_weights_manager && _weights_manager->are_weights_managed(b))
-    {
-        _reshape_rhs_kernel_managed->configure(compile_context, b, rhs_info);
-        reshaped_rhs = utils::cast::polymorphic_downcast<ICLTensor *>(_weights_manager->acquire(b, _reshape_rhs_kernel_managed.get()));
-    }
-    else
-    {
-        _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info);
-    }
-
-    // Configure and tune matrix multiply kernel
-    _mm_kernel->configure(compile_context, &_tmp_a, reshaped_rhs, c, output, alpha, beta, true, reshape_info, gemm_info.fp_mixed_precision(), gemm_info.activation_info());
-
-    CLScheduler::get().tune_kernel_static(*_mm_kernel);
-
-    // Allocate intermediate tensors
-    _tmp_a.allocator()->allocate();
-
-    if(!_reshape_b_only_on_first_run && use_mm_b)
-    {
-        _tmp_b.allocator()->allocate();
-    }
-}
-
-void CLGEMM::configure_reshaped_v2(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta,
-                                   const GEMMInfo &gemm_info)
-{
-    DataType           data_type               = a->info()->data_type();
-    bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
-    const unsigned int m                       = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
-    const unsigned int n                       = b->info()->dimension(0);
-    const unsigned int k                       = a->info()->dimension(0);
-    const unsigned int batch_size              = reinterpret_input_as_3d ? a->info()->dimension(3) : a->info()->dimension(2);
-    const int          depth_output_gemm3d     = gemm_info.depth_output_gemm3d();
-    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 = false;
-    kernel_info.broadcast_bias          = broadcast_bias;
-    kernel_info.activation_info         = gemm_info.activation_info();
-
-    // Set the target for the kernels
-    _reshape_lhs_kernel->set_target(gpu_target);
-    _mm_kernel->set_target(gpu_target);
-
-    const bool use_mm_b = (!_weights_manager || !_weights_manager->are_weights_managed(b));
-
-    // Manage intermediate buffers
-    _memory_group.manage(&_tmp_a);
-
-    if(!_reshape_b_only_on_first_run && use_mm_b)
-    {
-        _memory_group.manage(&_tmp_b);
-    }
-
-    // _tmp_a and _tmp_b will be auto configured in _interleave_kernel and in _transpose_kernel
-
-    GEMMLHSMatrixInfo lhs_info{};
-    GEMMRHSMatrixInfo rhs_info{};
-
-    // Pick up the GEMM configuration
-    std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }, kernel_info, a->info(), b->info(),
-                                                                    c == nullptr ? nullptr : c->info(), output->info(), gemm_info.reinterpret_input_as_3d());
-
-    _reshape_lhs_kernel->configure(compile_context, a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d());
-
-    ICLTensor *reshaped_rhs = &_tmp_b;
-    if(_weights_manager && _weights_manager->are_weights_managed(b))
-    {
-        _reshape_rhs_kernel_managed->configure(compile_context, b, rhs_info);
-        reshaped_rhs = utils::cast::polymorphic_downcast<ICLTensor *>(_weights_manager->acquire(b, _reshape_rhs_kernel_managed.get()));
-    }
-    else
-    {
-        _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info);
-    }
-
-    // Configure and tune matrix multiply kernel
-    _mm_reshaped_kernel->configure(compile_context, &_tmp_a, reshaped_rhs, c, output, alpha, beta, lhs_info, rhs_info, kernel_info);
-
-    // Allocate intermediate tensors
-    _tmp_a.allocator()->allocate();
-
-    if(!_reshape_b_only_on_first_run && use_mm_b)
-    {
-        _tmp_b.allocator()->allocate();
-    }
-}
-
-void CLGEMM::configure_reshaped_only_rhs(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta,
-                                         const GEMMInfo &gemm_info)
-{
-    DataType           data_type               = a->info()->data_type();
-    bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
-    const unsigned int m                       = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
-    const unsigned int n                       = b->info()->dimension(0);
-    const unsigned int k                       = a->info()->dimension(0);
-    const unsigned int batch_size              = reinterpret_input_as_3d ? a->info()->dimension(3) : a->info()->dimension(2);
-    const int          depth_output_gemm3d     = gemm_info.depth_output_gemm3d();
-    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);
-
-    const bool use_mm_b = (!_weights_manager || !_weights_manager->are_weights_managed(b));
-
-    // Manage intermediate buffers
-    if(!_reshape_b_only_on_first_run && use_mm_b)
-    {
-        _memory_group.manage(&_tmp_b);
-    }
-
-    GEMMLHSMatrixInfo lhs_info{};
-    GEMMRHSMatrixInfo rhs_info{};
-
-    // Pick up the GEMM configuration
-    std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }, kernel_info, a->info(), b->info(),
-                                                                             c == nullptr ? nullptr : c->info(), output->info());
-
-    ICLTensor *reshaped_rhs = &_tmp_b;
-    if(_weights_manager && _weights_manager->are_weights_managed(b))
-    {
-        _reshape_rhs_kernel_managed->configure(compile_context, b, rhs_info);
-        reshaped_rhs = utils::cast::polymorphic_downcast<ICLTensor *>(_weights_manager->acquire(b, _reshape_rhs_kernel_managed.get()));
-    }
-    else
-    {
-        _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info);
-    }
-
-    // Configure two variants of CLGEMMMatrixMultiplyReshapedOnlyRHSKernel (has_pad_y = false/true)
-    // During the prepare stage we check the padding requirement for the lhs and dst tensors. If they do not have
-    // pad y, we dispatch CLGEMMMatrixMultiplyReshapedOnlyRHSKernel with has_pad_y = false
-
-    // Configure matrix multiply kernel with no y padding support
-    kernel_info.has_pad_y = false;
-    _mm_reshaped_only_rhs_kernel->configure(compile_context, a, reshaped_rhs, c, output, alpha, beta, lhs_info, rhs_info, kernel_info);
-
-    // Configure matrix multiply kernel with y padding support
-    kernel_info.has_pad_y = true;
-    _mm_reshaped_only_rhs_fallback_kernel->configure(compile_context, a, reshaped_rhs, c, output, alpha, beta, lhs_info, rhs_info, kernel_info);
-
-    if(!_reshape_b_only_on_first_run && use_mm_b)
-    {
-        _tmp_b.allocator()->allocate();
-    }
-}
-
-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)
-{
-    ARM_COMPUTE_UNUSED(alpha);
-    ARM_COMPUTE_UNUSED(output);
-
-    // Get the GPU target
-    const GPUTarget    gpu_target              = CLScheduler::get().target();
-    bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
-    const unsigned int m                       = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
-    const unsigned int n                       = b->dimension(0);
-    const unsigned int k                       = a->dimension(0);
-    const int          depth_output_gemm3d     = gemm_info.depth_output_gemm3d();
-
-    const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d, gemm_info.broadcast_bias());
-
-    // 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()));
-
-    return Status{};
-}
-
-Status CLGEMM::validate_reshaped(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();
-    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();
-
-    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 = false;
-    kernel_info.broadcast_bias          = broadcast_bias;
-    kernel_info.activation_info         = gemm_info.activation_info();
-
-    GEMMLHSMatrixInfo lhs_info;
-    GEMMRHSMatrixInfo rhs_info;
-
-    // Pick up the GEMM configuration
-    // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails
-    const auto gemm_config = select_default_gemm_config_reshaped(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size });
-    lhs_info               = gemm_config.lhs_info;
-    rhs_info               = gemm_config.rhs_info;
-
-    auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, gemm_info.reinterpret_input_as_3d())));
-    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMReshapeLHSMatrixKernel::validate(a, &tmp_a_info, lhs_info, gemm_info.reinterpret_input_as_3d()));
-
-    auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
-    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMReshapeRHSMatrixKernel::validate(b, &tmp_b_info, rhs_info));
-
-    // Validate matrix multiply
-    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixMultiplyReshapedKernel::validate(&tmp_a_info, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info));
-
-    return Status{};
-}
-
-Status CLGEMM::validate_reshaped_only_rhs(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_b_info{};
-
-    // Get the GPU target
-    const GPUTarget    gpu_target              = CLScheduler::get().target();
-    const 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();
-
-    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();
-
-    GEMMLHSMatrixInfo lhs_info;
-    GEMMRHSMatrixInfo rhs_info;
-
-    // Pick up the GEMM configuration
-    // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails
-    const auto gemm_config = select_default_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size });
-    lhs_info               = gemm_config.lhs_info;
-    rhs_info               = gemm_config.rhs_info;
-
-    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
-    kernel_info.has_pad_y = false;
-    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info));
-
-    kernel_info.has_pad_y = true;
-    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info));
-
-    return Status{};
-}
-
 void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info)
 {
     configure(CLKernelLibrary::get().get_compile_context(), a, b, c, output, alpha, beta, gemm_info);
@@ -672,221 +74,56 @@ void CLGEMM::configure(const CLCompileContext &compile_context, const ICLTensor
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
 
-    // Perform validation step
-    ARM_COMPUTE_ERROR_THROW_ON(validate(a->info(), b->info(), c != nullptr ? c->info() : nullptr, output->info(), alpha, beta, gemm_info));
-
-    // Check if we need to reshape the matrix B only on the first run
-    _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
-    _is_prepared                 = gemm_info.retain_internal_weights();
-    _original_b                  = b;
-    _lhs                         = a;
-    _dst                         = output;
-
-    bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
-    const unsigned int m                       = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
-    const unsigned int n                       = b->info()->dimension(0);
-    const unsigned int k                       = a->info()->dimension(0);
-    const unsigned int batch_size              = reinterpret_input_as_3d ? a->info()->dimension(3) : a->info()->dimension(2);
-
-    // Select GEMMType
-    _gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery{ CLScheduler::get().target(), a->info()->data_type(), m, n, k, batch_size }, _reshape_b_only_on_first_run);
+    _impl->a   = a;
+    _impl->b   = b;
+    _impl->c   = c;
+    _impl->dst = output;
+    _impl->op  = std::make_unique<OperatorType>();
 
-    const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr);
+    _impl->op->configure(compile_context, a->info(), b->info(), c != nullptr ? c->info() : nullptr, output->info(), alpha, beta, gemm_info);
+    _impl->aux_mem_req = _impl->op->workspace();
 
-    const ICLTensor *c_to_use = fuse_add_c ? c : nullptr;
-
-    switch(_gemm_kernel_type)
-    {
-        case CLGEMMKernelType::NATIVE_V1:
-        {
-            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);
-            break;
-        }
-        case CLGEMMKernelType::RESHAPED:
-        {
-            configure_reshaped_v2(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info);
-            break;
-        }
-        case CLGEMMKernelType::RESHAPED_ONLY_RHS:
-        {
-            configure_reshaped_only_rhs(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info);
-            break;
-        }
-        default:
-        {
-            ARM_COMPUTE_ERROR("GEMMType not supported");
-        }
-    }
+    // Manage/allocate auxilairy tensors
+    _impl->run_pack          = { { ACL_SRC_0, _impl->a }, { ACL_SRC_2, _impl->c }, { ACL_DST, _impl->dst } };
+    _impl->prep_pack         = { { ACL_SRC_1, _impl->b } };
+    _impl->workspace_tensors = manage_workspace<CLTensor>(_impl->op->workspace(), _impl->memory_group, _impl->run_pack, _impl->prep_pack);
 }
 
 Status CLGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info)
 {
-    // Get the GPU target
-    bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
-    const unsigned int m                       = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
-    const unsigned int n                       = b->dimension(0);
-    const unsigned int k                       = a->dimension(0);
-    const unsigned int batch_size              = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2);
-
-    // Select GEMMType
-    CLGEMMKernelType gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery
-    {
-        CLScheduler::get().target(), a->data_type(), m, n, k, batch_size,
-    },
-    gemm_info.reshape_b_only_on_first_run());
-
-    const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr);
-
-    const ITensorInfo *c_to_use = fuse_add_c ? c : nullptr;
-
-    switch(gemm_kernel_type)
-    {
-        case CLGEMMKernelType::NATIVE_V1:
-        {
-            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));
-            break;
-        }
-        case CLGEMMKernelType::RESHAPED:
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped(a, b, c_to_use, output, alpha, beta, gemm_info));
-            break;
-        }
-        case CLGEMMKernelType::RESHAPED_ONLY_RHS:
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped_only_rhs(a, b, c_to_use, output, alpha, beta, gemm_info));
-            break;
-        }
-        default:
-        {
-            ARM_COMPUTE_RETURN_ERROR_MSG("GEMMType not supported");
-        }
-    }
-
-    return Status{};
+    return OperatorType::validate(a, b, c, output, alpha, beta, gemm_info);
 }
 
 void CLGEMM::run()
 {
     prepare();
-    MemoryGroupResourceScope scope_mg(_memory_group);
-
-    // Run matrix multiply kernel
-    switch(_gemm_kernel_type)
-    {
-        case CLGEMMKernelType::NATIVE_V1:
-        {
-            CLScheduler::get().enqueue(*_mm_kernel, true);
-            break;
-        }
-        case CLGEMMKernelType::RESHAPED_V1:
-        {
-            // Run interleave kernel
-            CLScheduler::get().enqueue(*_reshape_lhs_kernel, false);
 
-            if(!_reshape_b_only_on_first_run)
-            {
-                // Run transpose kernel
-                if(_weights_manager && _weights_manager->are_weights_managed(_original_b))
-                {
-                    _weights_manager->run(_original_b, _reshape_rhs_kernel_managed.get());
-                }
-                else
-                {
-                    CLScheduler::get().enqueue(*_reshape_rhs_kernel, false);
-                }
-            }
+    MemoryGroupResourceScope scope_mg(_impl->memory_group);
 
-            CLScheduler::get().enqueue(*_mm_kernel, true);
-            break;
-        }
-        case CLGEMMKernelType::RESHAPED:
-        {
-            // Run interleave kernel
-            CLScheduler::get().enqueue(*_reshape_lhs_kernel, false);
-
-            if(!_reshape_b_only_on_first_run)
-            {
-                // Run transpose kernel
-                if(_weights_manager && _weights_manager->are_weights_managed(_original_b))
-                {
-                    _weights_manager->run(_original_b, _reshape_rhs_kernel_managed.get());
-                }
-                else
-                {
-                    CLScheduler::get().enqueue(*_reshape_rhs_kernel, false);
-                }
-            }
-
-            CLScheduler::get().enqueue(*_mm_reshaped_kernel, true);
-            break;
-        }
-        case CLGEMMKernelType::RESHAPED_ONLY_RHS:
-        {
-            if(!_reshape_b_only_on_first_run)
-            {
-                // Run transpose kernel
-                if(_weights_manager && _weights_manager->are_weights_managed(_original_b))
-                {
-                    _weights_manager->run(_original_b, _reshape_rhs_kernel_managed.get());
-                }
-                else
-                {
-                    CLScheduler::get().enqueue(*_reshape_rhs_kernel, false);
-                }
-            }
-            // In case of RESHAPED_ONLY_RHS, we need to check the padding requirement
-            // Check if the lhs or dst tensors have padding
-            const unsigned int cross_plane_pad_lhs = _lhs->info()->padding().top + _lhs->info()->padding().bottom;
-            const unsigned int cross_plane_pad_dst = _dst->info()->padding().top + _dst->info()->padding().bottom;
-
-            bool has_pad_y = (cross_plane_pad_lhs != 0) || (cross_plane_pad_dst != 0);
-            if(has_pad_y)
-            {
-                CLScheduler::get().enqueue(*_mm_reshaped_only_rhs_fallback_kernel, true);
-            }
-            else
-            {
-                CLScheduler::get().enqueue(*_mm_reshaped_only_rhs_kernel, true);
-            }
-            break;
-        }
-        default:
-        {
-            ARM_COMPUTE_ERROR("GEMMType not supported");
-        }
-    }
+    ARM_COMPUTE_ERROR_ON_NULLPTR(_impl->a, _impl->b, _impl->dst);
+    _impl->op->run(_impl->run_pack);
 }
 
 void CLGEMM::prepare()
 {
-    if(!_is_prepared)
+    if(!_impl->_is_prepared)
     {
-        if(_gemm_kernel_type != CLGEMMKernelType::NATIVE_V1 && _reshape_b_only_on_first_run)
+        _impl->op->prepare(_impl->prep_pack);
+
+        auto has_reshape = std::find_if(_impl->aux_mem_req.begin(),
+                                        _impl->aux_mem_req.end(),
+                                        [](const MemoryInfo & m) -> bool { return m.lifetime == MemoryLifetime::Persistent; });
+
+        if(has_reshape != std::end(_impl->aux_mem_req))
+        {
+            _impl->b->mark_as_unused();
+        }
+        else
         {
-            if(_weights_manager && _weights_manager->are_weights_managed(_original_b))
-            {
-                _weights_manager->run(_original_b, _reshape_rhs_kernel_managed.get());
-            }
-            else
-            {
-                // Run transpose kernel and mark original weights tensor as unused
-                _tmp_b.allocator()->allocate();
-                CLScheduler::get().enqueue(*_reshape_rhs_kernel, false);
-                _original_b->mark_as_unused();
-            }
+            // Pack the B matrix to be used as the underlying GEMM performs no reshapes
+            _impl->run_pack.add_const_tensor(ACL_SRC_1, _impl->b);
         }
-        CLScheduler::get().queue().finish();
-        _is_prepared = true;
+        _impl->_is_prepared = true;
     }
 }
 } // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index f37f06b0ff..5dc7556b2f 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,11 +37,6 @@
 #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"
-#include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 #include "src/core/CL/kernels/CLIm2ColKernel.h"
 #include "src/core/CL/kernels/CLWeightsReshapeKernel.h"
 #include "src/core/helpers/AutoConfiguration.h"
diff --git a/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
index a040e9d38e..7a01018f59 100644
--- a/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -36,11 +36,6 @@
 #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"
-#include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 #include "src/core/CL/kernels/CLIm2ColKernel.h"
 #include "src/core/CL/kernels/CLWeightsReshapeKernel.h"
 
diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index 5a9ff7990f..099a2c980f 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -40,7 +40,7 @@
 #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/CLGEMMReshapeRHSMatrixKernel.h"
+#include "src/core/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.h"
 #include "utils/TypePrinter.h"
@@ -127,7 +127,7 @@ inline bool validate_lhs_rhs_info_reshaped_only_rhs(const GEMMLHSMatrixInfo &lhs
     TensorInfo tmp_b_info{};
     // Validate reshape RHS kernel
     auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
-    if(!bool(CLGEMMReshapeRHSMatrixKernel::validate(b, &tmp_b_info, rhs_info)))
+    if(!bool(opencl::kernels::ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)))
     {
         return false;
     }
@@ -192,7 +192,7 @@ CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemo
       _weights_to_qasymm8(std::make_unique<CLDepthConvertLayerKernel>()),
       _mm_native_kernel(std::make_unique<CLGEMMLowpMatrixMultiplyNativeKernel>()),
       _mm_reshaped_only_rhs_kernel(std::make_unique<CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel>()),
-      _mtx_b_reshape_kernel(std::make_unique<CLGEMMReshapeRHSMatrixKernel>()),
+      _mtx_b_reshape_kernel(std::make_unique<opencl::kernels::ClGemmReshapeRhsMatrixKernel>()),
       _mtx_a_reduction_kernel(std::make_unique<CLGEMMLowpMatrixAReductionKernel>()),
       _mtx_b_reduction_kernel(std::make_unique<CLGEMMLowpMatrixBReductionKernel>()),
       _offset_contribution_kernel(std::make_unique<CLGEMMLowpOffsetContributionKernel>()),
@@ -292,7 +292,7 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const CLCompileContext &compile_con
                                                                                  a->info(), _convert_to_qasymm8 ? _qasymm8_weights.info() : b->info(), output->info());
 
         // Configure reshape RHS kernel
-        _mtx_b_reshape_kernel->configure(compile_context, _convert_to_qasymm8 ? &_qasymm8_weights : b, &_tmp_b, rhs_info);
+        _mtx_b_reshape_kernel->configure(compile_context, _convert_to_qasymm8 ? _qasymm8_weights.info() : b->info(), _tmp_b.info(), rhs_info);
     }
 
     // Using default reduction info
@@ -496,7 +496,7 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
 
         // Validate reshape RHS kernel
         auto_init_if_empty(tmp_b_info, weights_info.clone()->set_tensor_shape(compute_rhs_reshaped_shape(weights_info, rhs_info)));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMReshapeRHSMatrixKernel::validate(&weights_info, &tmp_b_info, rhs_info));
+        ARM_COMPUTE_RETURN_ON_ERROR(opencl::kernels::ClGemmReshapeRhsMatrixKernel::validate(&weights_info, &tmp_b_info, rhs_info));
     }
 
     TensorInfo info_vector_sum_col{};
@@ -634,6 +634,9 @@ void CLGEMMLowpMatrixMultiplyCore::run()
         if(!_reshape_b_only_on_first_run)
         {
             // Run reshape matrix B
+            ITensorPack mtx_b_pack;
+            mtx_b_pack.add_const_tensor(TensorType::ACL_SRC, _convert_to_qasymm8 ? &_qasymm8_weights : _original_b);
+            mtx_b_pack.add_tensor(TensorType::ACL_DST, &_tmp_b);
             CLScheduler::get().enqueue(*_mtx_b_reshape_kernel, false);
         }
     }
@@ -687,7 +690,10 @@ void CLGEMMLowpMatrixMultiplyCore::prepare()
 
             // Run reshape kernel and mark original weights tensor as unused
             _tmp_b.allocator()->allocate();
-            CLScheduler::get().enqueue(*_mtx_b_reshape_kernel, false);
+            ITensorPack mtx_b_pack;
+            mtx_b_pack.add_const_tensor(TensorType::ACL_SRC, _convert_to_qasymm8 ? &_qasymm8_weights : _original_b);
+            mtx_b_pack.add_tensor(TensorType::ACL_DST, &_tmp_b);
+            CLScheduler::get().enqueue_op(*_mtx_b_reshape_kernel, mtx_b_pack, false);
             _original_b->mark_as_unused();
         }
 
diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp
index 05d459c899..146ac8f619 100644
--- a/src/runtime/CL/functions/CLLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayer.cpp
@@ -36,11 +36,6 @@
 #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"
-#include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 #include "src/core/gpu/cl/kernels/ClTransposeKernel.h"
 
 namespace arm_compute
diff --git a/src/runtime/CL/functions/CLLSTMLayerQuantized.cpp b/src/runtime/CL/functions/CLLSTMLayerQuantized.cpp
index 46062387e7..69974424c9 100644
--- a/src/runtime/CL/functions/CLLSTMLayerQuantized.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayerQuantized.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -34,7 +34,6 @@
 #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/CLGEMMReshapeRHSMatrixKernel.h"
 #include "src/core/helpers/AutoConfiguration.h"
 
 #include <memory>
diff --git a/src/runtime/CL/functions/CLQLSTMLayer.cpp b/src/runtime/CL/functions/CLQLSTMLayer.cpp
index e7a0e5765e..7b6ec8f5c8 100644
--- a/src/runtime/CL/functions/CLQLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLQLSTMLayer.cpp
@@ -37,7 +37,6 @@
 #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/CLGEMMReshapeRHSMatrixKernel.h"
 #include "src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h"
 #include "src/core/helpers/WindowHelpers.h"
 
diff --git a/src/runtime/CL/functions/CLRNNLayer.cpp b/src/runtime/CL/functions/CLRNNLayer.cpp
index 967f4aa41b..45ced35782 100644
--- a/src/runtime/CL/functions/CLRNNLayer.cpp
+++ b/src/runtime/CL/functions/CLRNNLayer.cpp
@@ -35,11 +35,6 @@
 #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"
-#include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 
 namespace arm_compute
 {
diff --git a/src/runtime/CL/functions/CLSoftmaxLayer.cpp b/src/runtime/CL/functions/CLSoftmaxLayer.cpp
index e47537bd31..fe45f65beb 100644
--- a/src/runtime/CL/functions/CLSoftmaxLayer.cpp
+++ b/src/runtime/CL/functions/CLSoftmaxLayer.cpp
@@ -43,7 +43,7 @@ struct CLSoftmaxLayerGeneric<IS_LOG>::Impl
     ICLTensor                    *dst{ nullptr };
     std::unique_ptr<OperatorType> op{ nullptr };
     MemoryGroup                   memory_group{};
-    std::vector<std::pair<TensorType, std::unique_ptr<CLTensor>>> workspace_tensors{};
+    std::vector<std::pair<int, std::unique_ptr<CLTensor>>> workspace_tensors{};
 };
 
 template <bool IS_LOG>
@@ -88,14 +88,14 @@ void           CLSoftmaxLayerGeneric<IS_LOG>::allocate_workspace()
     std::for_each(memory_requirements.begin(), memory_requirements.end(), [this](const experimental::MemoryInfo & memory_info)
     {
         auto tensor_info = TensorInfo{ TensorShape(memory_info.size), 1, DataType::U8 };
-        _impl->workspace_tensors.emplace_back(memory_info.type, std::make_unique<CLTensor>());
+        _impl->workspace_tensors.emplace_back(memory_info.slot, std::make_unique<CLTensor>());
         auto tensor = _impl->workspace_tensors.back().second.get();
         ARM_COMPUTE_ERROR_ON_NULLPTR(tensor);
         tensor->allocator()->init(tensor_info);
         _impl->memory_group.manage(tensor);
     });
 
-    std::for_each(_impl->workspace_tensors.begin(), _impl->workspace_tensors.end(), [](std::pair<TensorType, std::unique_ptr<CLTensor>> &wt)
+    std::for_each(_impl->workspace_tensors.begin(), _impl->workspace_tensors.end(), [](std::pair<int, std::unique_ptr<CLTensor>> &wt)
     {
         auto tensor = wt.second.get();
         tensor->allocator()->allocate();
@@ -114,7 +114,7 @@ void           CLSoftmaxLayerGeneric<IS_LOG>::run()
     pack.add_tensor(TensorType::ACL_SRC, _impl->src);
     pack.add_tensor(TensorType::ACL_DST, _impl->dst);
 
-    std::for_each(_impl->workspace_tensors.begin(), _impl->workspace_tensors.end(), [&pack](std::pair<TensorType, std::unique_ptr<CLTensor>> &wt)
+    std::for_each(_impl->workspace_tensors.begin(), _impl->workspace_tensors.end(), [&pack](std::pair<int, std::unique_ptr<CLTensor>> &wt)
     {
         auto tensor = wt.second.get();
         ARM_COMPUTE_ERROR_ON_NULLPTR(tensor);
diff --git a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp
index 321466f05f..6b8b00414a 100644
--- a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,11 +29,6 @@
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "src/core/CL/kernels/CLFillBorderKernel.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"
-#include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 #include "src/core/CL/kernels/CLWinogradFilterTransformKernel.h"
 #include "src/core/CL/kernels/CLWinogradOutputTransformKernel.h"
 
diff --git a/src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp b/src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp
index 5ac25a9a20..390bb97665 100644
--- a/src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp
+++ b/src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020 Arm Limited.
+ * Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,7 +25,7 @@
 
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "src/core/CL/gemm/CLGEMMHelpers.h"
+#include "src/core/gpu/cl/kernels/gemm/ClGemmHelpers.h"
 
 #include <map>
 #include <utility>
diff --git a/src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp b/src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp
index 88b6060e12..b799de6967 100644
--- a/src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp
+++ b/src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020 Arm Limited.
+ * Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,7 +26,7 @@
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/GPUTarget.h"
-#include "src/core/CL/gemm/CLGEMMHelpers.h"
+#include "src/core/gpu/cl/kernels/gemm/ClGemmHelpers.h"
 
 #include <map>
 #include <utility>
diff --git a/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp b/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
index 0f754276c7..982748810d 100644
--- a/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
+++ b/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
@@ -25,7 +25,7 @@
 
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "src/core/CL/gemm/CLGEMMHelpers.h"
+#include "src/core/gpu/cl/kernels/gemm/ClGemmHelpers.h"
 
 #include <map>
 #include <utility>
diff --git a/src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp b/src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp
index 489be356d9..b8437487f8 100644
--- a/src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp
+++ b/src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp
@@ -27,11 +27,11 @@
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "arm_compute/runtime/CL/ICLGEMMKernelSelection.h"
-#include "src/core/CL/ICLGEMMKernelConfiguration.h"
-#include "src/core/CL/gemm/CLGEMMHelpers.h"
-#include "src/core/CL/gemm/native/CLGEMMNativeKernelConfiguration.h"
-#include "src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfiguration.h"
-#include "src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfiguration.h"
+#include "src/core/gpu/cl/kernels/gemm/ClGemmHelpers.h"
+#include "src/core/gpu/cl/kernels/gemm/IClGemmKernelConfig.h"
+#include "src/core/gpu/cl/kernels/gemm/native/ClGemmNativeKernelConfig.h"
+#include "src/core/gpu/cl/kernels/gemm/reshaped/ClGemmReshapedKernelConfig.h"
+#include "src/core/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmReshapedOnlyRhsKernelConfig.h"
 #include "src/runtime/CL/gemm/CLGEMMKernelSelection.h"
 #include "src/runtime/CL/mlgo/MLGOHeuristics.h"
 #include "src/runtime/CL/mlgo/Utils.h"
@@ -43,6 +43,8 @@ namespace cl_gemm
 {
 namespace auto_heuristics
 {
+using namespace arm_compute::opencl::kernels::gemm;
+
 GEMMTypeResult select_mlgo_gemm_kernel(const CommonQuery &query, bool reshape_b_only_on_first_run)
 {
     ARM_COMPUTE_UNUSED(reshape_b_only_on_first_run);
@@ -83,9 +85,9 @@ GEMMTypeResult select_default_gemm_kernel(const CommonQuery &query, bool reshape
 
 GEMMConfigResult select_default_gemm_config_reshaped_only_rhs(const CommonQuery &query)
 {
-    GEMMLHSMatrixInfo                           lhs_info;
-    GEMMRHSMatrixInfo                           rhs_info;
-    std::unique_ptr<ICLGEMMKernelConfiguration> gemm_config = CLGEMMReshapedOnlyRHSKernelConfigurationFactory::create(query.gpu_target);
+    GEMMLHSMatrixInfo                    lhs_info;
+    GEMMRHSMatrixInfo                    rhs_info;
+    std::unique_ptr<IClGemmKernelConfig> gemm_config = ClGemmReshapedOnlyRhsKernelConfigurationFactory::create(query.gpu_target);
     ARM_COMPUTE_ERROR_ON_NULLPTR(gemm_config.get());
     std::tie(lhs_info, rhs_info) = gemm_config->configure(query.m, query.n, query.k, query.b, query.data_type);
     return GEMMConfigResult{ true, lhs_info, rhs_info };
@@ -118,9 +120,9 @@ GEMMConfigResult select_mlgo_gemm_config_reshaped_only_rhs(const CommonQuery &qu
 
 GEMMConfigResult select_default_gemm_config_reshaped(const CommonQuery &query)
 {
-    GEMMLHSMatrixInfo                           lhs_info;
-    GEMMRHSMatrixInfo                           rhs_info;
-    std::unique_ptr<ICLGEMMKernelConfiguration> gemm_config = CLGEMMReshapedKernelConfigurationFactory::create(query.gpu_target);
+    GEMMLHSMatrixInfo                    lhs_info;
+    GEMMRHSMatrixInfo                    rhs_info;
+    std::unique_ptr<IClGemmKernelConfig> gemm_config = ClGemmReshapedKernelConfigurationFactory::create(query.gpu_target);
     ARM_COMPUTE_ERROR_ON_NULLPTR(gemm_config.get());
     std::tie(lhs_info, rhs_info) = gemm_config->configure(query.m, query.n, query.k, query.b, query.data_type);
     return GEMMConfigResult{ true, lhs_info, rhs_info };
@@ -152,9 +154,9 @@ GEMMConfigResult select_mlgo_gemm_config_reshaped(const CommonQuery &query)
 
 GEMMConfigResult select_default_gemm_config_native(const CommonQuery &query)
 {
-    GEMMLHSMatrixInfo                           lhs_info;
-    GEMMRHSMatrixInfo                           rhs_info;
-    std::unique_ptr<ICLGEMMKernelConfiguration> gemm_config = CLGEMMNativeKernelConfigurationFactory::create(query.gpu_target);
+    GEMMLHSMatrixInfo                    lhs_info;
+    GEMMRHSMatrixInfo                    rhs_info;
+    std::unique_ptr<IClGemmKernelConfig> gemm_config = ClGemmNativeKernelConfigurationFactory::create(query.gpu_target);
     ARM_COMPUTE_ERROR_ON_NULLPTR(gemm_config.get());
     std::tie(lhs_info, rhs_info) = gemm_config->configure(query.m, query.n, query.k, query.b, query.data_type);
     return GEMMConfigResult{ true, lhs_info, rhs_info };
@@ -175,7 +177,7 @@ GEMMConfigResult select_mlgo_gemm_config_native(const CommonQuery &query)
     {
         ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("MLGOHeuristics query returns gemm config: %s.", to_string(config).c_str());
         // Setting irrelevant unsigned int parameters to 1 and bool parameters to false as they do no matter
-        std::tie(lhs_info, rhs_info) = configure_lhs_rhs_info(query.m, query.n, config.m0, config.n0, config.k0, 1, 1, false, false, false, false, false);
+        std::tie(lhs_info, rhs_info) = opencl::kernels::gemm::configure_lhs_rhs_info(query.m, query.n, config.m0, config.n0, config.k0, 1, 1, false, false, false, false, false);
     }
     else
     {
diff --git a/src/runtime/ITensorAllocator.cpp b/src/runtime/ITensorAllocator.cpp
index ae648d4dd2..fe3d2804cb 100644
--- a/src/runtime/ITensorAllocator.cpp
+++ b/src/runtime/ITensorAllocator.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -30,25 +30,27 @@
 
 using namespace arm_compute;
 
-ITensorAllocator::ITensorAllocator()
-    : _info(), _alignment(0)
+void ITensorAllocator::init(const TensorInfo &input, size_t alignment)
 {
+    _info_owned    = input;
+    _info_external = nullptr;
+    _alignment     = alignment;
 }
 
-void ITensorAllocator::init(const TensorInfo &input, size_t alignment)
+void ITensorAllocator::soft_init(TensorInfo &input, size_t alignment)
 {
-    _info      = input;
-    _alignment = alignment;
+    _info_external = &input;
+    _alignment     = alignment;
 }
 
 TensorInfo &ITensorAllocator::info()
 {
-    return _info;
+    return (_info_external != nullptr) ? *_info_external : _info_owned;
 }
 
 const TensorInfo &ITensorAllocator::info() const
 {
-    return _info;
+    return (_info_external != nullptr) ? *_info_external : _info_owned;
 }
 
 size_t ITensorAllocator::alignment() const
diff --git a/src/runtime/gpu/cl/operators/ClGemm.cpp b/src/runtime/gpu/cl/operators/ClGemm.cpp
new file mode 100644
index 0000000000..fcbc6d5fba
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClGemm.cpp
@@ -0,0 +1,754 @@
+/*
+ * 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/runtime/gpu/cl/operators/ClGemm.h"
+
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/GPUTarget.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/KernelDescriptors.h"
+#include "arm_compute/core/Log.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+#include "arm_compute/runtime/ITensorAllocator.h"
+#include "src/core/gpu/cl/IClKernel.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/MemoryHelpers.h"
+#include "src/core/utils/helpers/float_ops.h"
+#include "src/runtime/CL/gemm/CLGEMMKernelSelection.h"
+#include "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.h"
+#include "src/runtime/gpu/cl/utils/ClAuxTensorHandler.h"
+
+#include "support/Cast.h"
+#include "utils/TypePrinter.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+using namespace arm_compute::misc::shape_calculator;
+using namespace arm_compute::cl_gemm;
+using namespace arm_compute::experimental;
+using namespace arm_compute::utils::cast;
+using namespace arm_compute::opencl::kernels;
+
+namespace
+{
+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;
+        }
+    }
+}
+//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)
+{
+    auto gemm_kernel = auto_heuristics::select_mlgo_gemm_kernel(query, reshape_b_only_on_first_run);
+    if(bool(gemm_kernel))
+    {
+        if(validate_gemm_kernel(gemm_kernel.gemm_type))
+        {
+            ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from mlgo heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str());
+            return gemm_kernel.gemm_type;
+        }
+    }
+    gemm_kernel = auto_heuristics::select_default_gemm_kernel(query, reshape_b_only_on_first_run);
+    ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from default heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str());
+    return gemm_kernel.gemm_type;
+}
+// Validate lhs_info and rhs_info for reshaped only rhs kernel
+inline bool validate_lhs_rhs_info_reshaped_only_rhs(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c,
+                                                    const ITensorInfo *output, GEMMKernelInfo gemm_kernel_info)
+{
+    // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel
+    TensorInfo tmp_b_info{};
+    // Validate reshape RHS kernel
+    auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
+    if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)))
+    {
+        return false;
+    }
+    // Validate mm kernel
+    gemm_kernel_info.lhs_info  = lhs_info;
+    gemm_kernel_info.rhs_info  = rhs_info;
+    gemm_kernel_info.has_pad_y = false;
+    if(!bool(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info)))
+    {
+        return false;
+    }
+    gemm_kernel_info.has_pad_y = true;
+    if(!bool(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info)))
+    {
+        return false;
+    }
+    return true;
+}
+
+//Automatically select between mlgo (prioritized) and default heuristics for reshaped only rhs kernel configs
+inline std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery query, GEMMKernelInfo kernel_info, const ITensorInfo *a,
+                                                                                                 const ITensorInfo *b,
+                                                                                                 const ITensorInfo *c, const ITensorInfo *output)
+{
+    auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(query);
+    if(config)
+    {
+        if(validate_lhs_rhs_info_reshaped_only_rhs(config.lhs_info, config.rhs_info, a, b, c, output, kernel_info))
+        {
+            ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
+            return { config.lhs_info, config.rhs_info };
+        }
+    }
+    config = auto_heuristics::select_default_gemm_config_reshaped_only_rhs(query);
+    ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
+    return { config.lhs_info, config.rhs_info };
+}
+
+// Validate lhs_info and rhs_info for reshaped kernel
+inline bool validate_lhs_rhs_info_reshaped(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c,
+                                           const ITensorInfo *output, GEMMKernelInfo gemm_kernel_info, bool reinterpret_input_as_3d)
+{
+    // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped kernel
+    TensorInfo tmp_a_info{};
+    TensorInfo tmp_b_info{};
+
+    // Validate reshape LHS kernel
+    auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, reinterpret_input_as_3d)));
+    if(!bool(ClGemmReshapeLhsMatrixKernel::validate(a, &tmp_a_info, lhs_info, reinterpret_input_as_3d)))
+    {
+        return false;
+    }
+
+    // Validate reshape RHS kernel
+    auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
+    if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)))
+    {
+        return false;
+    }
+    // Validate mm kernel
+    gemm_kernel_info.lhs_info = lhs_info;
+    gemm_kernel_info.rhs_info = rhs_info;
+    if(!bool(ClGemmMatrixMultiplyReshapedKernel::validate(&tmp_a_info, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info)))
+    {
+        return false;
+    }
+    return true;
+}
+
+//Automatically select between mlgo (prioritized) and default heuristics for reshaped kernel configs
+inline std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped(auto_heuristics::CommonQuery query, GEMMKernelInfo kernel_info, const ITensorInfo *a, const ITensorInfo *b,
+                                                                                        const ITensorInfo *c, const ITensorInfo *output, bool reinterpret_input_as_3d)
+{
+    auto config = auto_heuristics::select_mlgo_gemm_config_reshaped(query);
+    if(config)
+    {
+        if(validate_lhs_rhs_info_reshaped(config.lhs_info, config.rhs_info, a, b, c, output, kernel_info, reinterpret_input_as_3d))
+        {
+            ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
+            return { config.lhs_info, config.rhs_info };
+        }
+    }
+    config = auto_heuristics::select_default_gemm_config_reshaped(query);
+    ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
+    return { config.lhs_info, config.rhs_info };
+}
+} // namespace
+
+ClGemm::ClGemm()
+    : _mm_kernel(std::make_unique<ClGemmMatrixMultiplyKernel>()),
+      _reshape_lhs_kernel(std::make_unique<ClGemmReshapeLhsMatrixKernel>()),
+      _reshape_rhs_kernel(std::make_unique<ClGemmReshapeRhsMatrixKernel>()),
+      _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),
+      _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)
+{
+    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();
+
+    // Set the target for the kernels
+    _mm_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());
+
+    // 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);
+}
+
+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)
+{
+    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 = false;
+    kernel_info.broadcast_bias          = broadcast_bias;
+    kernel_info.activation_info         = gemm_info.activation_info();
+
+    // Set the target for the kernels
+    _reshape_lhs_kernel->set_target(gpu_target);
+    _mm_kernel->set_target(gpu_target);
+
+    GEMMLHSMatrixInfo lhs_info{};
+    GEMMRHSMatrixInfo rhs_info{};
+
+    // Pick up the GEMM configuration
+    std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }, kernel_info, a, b,
+                                                                    c, output, gemm_info.reinterpret_input_as_3d());
+
+    _reshape_lhs_kernel->configure(compile_context, a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d());
+    _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info);
+
+    // Configure and tune matrix multiply kernel
+    _mm_reshaped_kernel->configure(compile_context, &_tmp_a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info);
+
+    // 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_only_rhs(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();
+    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);
+
+    GEMMLHSMatrixInfo lhs_info{};
+    GEMMRHSMatrixInfo rhs_info{};
+
+    // Pick up the GEMM configuration
+    std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }, kernel_info, a, b, c, output);
+
+    // Transpose matrix
+    _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info);
+
+    // Configure two variants of CLGEMMMatrixMultiplyReshapedOnlyRHSKernel (has_pad_y = false/true)
+    // During the prepare stage we check the padding requirement for the lhs and dst tensors. If they do not have
+    // pad y, we dispatch CLGEMMMatrixMultiplyReshapedOnlyRHSKernel with has_pad_y = false
+
+    // Configure matrix multiply kernel with no y padding support
+    kernel_info.has_pad_y = false;
+    _mm_reshaped_only_rhs_kernel->configure(compile_context, a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info);
+
+    // Configure matrix multiply kernel with y padding support
+    kernel_info.has_pad_y = true;
+    _mm_reshaped_only_rhs_fallback_kernel->configure(compile_context, a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info);
+
+    // Request memory for RHS reshape matrix
+    _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)
+{
+    ARM_COMPUTE_UNUSED(alpha);
+    ARM_COMPUTE_UNUSED(output);
+
+    // Get the GPU target
+    const GPUTarget    gpu_target              = CLScheduler::get().target();
+    bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
+    const unsigned int m                       = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
+    const unsigned int n                       = b->dimension(0);
+    const unsigned int k                       = a->dimension(0);
+    const int          depth_output_gemm3d     = gemm_info.depth_output_gemm3d();
+
+    const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d, gemm_info.broadcast_bias());
+
+    // 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()));
+
+    return Status{};
+}
+
+Status ClGemm::validate_reshaped(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();
+    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();
+
+    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 = false;
+    kernel_info.broadcast_bias          = broadcast_bias;
+    kernel_info.activation_info         = gemm_info.activation_info();
+
+    GEMMLHSMatrixInfo lhs_info;
+    GEMMRHSMatrixInfo rhs_info;
+
+    // Pick up the GEMM configuration
+    // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails
+    const auto gemm_config = select_default_gemm_config_reshaped(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size });
+    lhs_info               = gemm_config.lhs_info;
+    rhs_info               = gemm_config.rhs_info;
+
+    auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, gemm_info.reinterpret_input_as_3d())));
+    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeLhsMatrixKernel::validate(a, &tmp_a_info, lhs_info, gemm_info.reinterpret_input_as_3d()));
+
+    auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
+    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info));
+
+    // Validate matrix multiply
+    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedKernel::validate(&tmp_a_info, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info));
+
+    return Status{};
+}
+
+Status ClGemm::validate_reshaped_only_rhs(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_b_info{};
+
+    // Get the GPU target
+    const GPUTarget    gpu_target              = CLScheduler::get().target();
+    const 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();
+
+    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();
+
+    GEMMLHSMatrixInfo lhs_info;
+    GEMMRHSMatrixInfo rhs_info;
+
+    // Pick up the GEMM configuration
+    // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails
+    const auto gemm_config = select_default_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size });
+    lhs_info               = gemm_config.lhs_info;
+    rhs_info               = gemm_config.rhs_info;
+
+    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
+    kernel_info.has_pad_y = false;
+    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info));
+
+    kernel_info.has_pad_y = true;
+    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info));
+
+    return Status{};
+}
+
+void ClGemm::configure(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate(a, b, c, output, alpha, beta, gemm_info));
+
+    // Check if we need to reshape the matrix B only on the first run
+    _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
+
+    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);
+
+    // Select GEMMType
+    _gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery{ CLScheduler::get().target(), a->data_type(), m, n, k, batch_size }, _reshape_b_only_on_first_run);
+
+    const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr);
+
+    ITensorInfo *c_to_use = fuse_add_c ? c : nullptr;
+
+    switch(_gemm_kernel_type)
+    {
+        case CLGEMMKernelType::NATIVE_V1:
+        {
+            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);
+            break;
+        }
+        case CLGEMMKernelType::RESHAPED:
+        {
+            configure_reshaped_v2(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info);
+            break;
+        }
+        case CLGEMMKernelType::RESHAPED_ONLY_RHS:
+        {
+            configure_reshaped_only_rhs(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info);
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("GEMMType not supported");
+        }
+    }
+}
+
+Status ClGemm::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info)
+{
+    // Get the GPU target
+    bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
+    const unsigned int m                       = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
+    const unsigned int n                       = b->dimension(0);
+    const unsigned int k                       = a->dimension(0);
+    const unsigned int batch_size              = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2);
+
+    // Select GEMMType
+    CLGEMMKernelType gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery
+    {
+        CLScheduler::get().target(), a->data_type(), m, n, k, batch_size,
+    },
+    gemm_info.reshape_b_only_on_first_run());
+
+    const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr);
+
+    const ITensorInfo *c_to_use = fuse_add_c ? c : nullptr;
+
+    switch(gemm_kernel_type)
+    {
+        case CLGEMMKernelType::NATIVE_V1:
+        {
+            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));
+            break;
+        }
+        case CLGEMMKernelType::RESHAPED:
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped(a, b, c_to_use, output, alpha, beta, gemm_info));
+            break;
+        }
+        case CLGEMMKernelType::RESHAPED_ONLY_RHS:
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped_only_rhs(a, b, c_to_use, output, alpha, beta, gemm_info));
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_RETURN_ERROR_MSG("GEMMType not supported");
+        }
+    }
+
+    return Status{};
+}
+
+void ClGemm::run(ITensorPack &tensors)
+{
+    const ITensor *lhs  = tensors.get_const_tensor(ACL_SRC_0);
+    const ITensor *rhs  = tensors.get_const_tensor(ACL_SRC_1);
+    const ITensor *src2 = tensors.get_const_tensor(ACL_SRC_2);
+    ITensor       *dst  = tensors.get_tensor(ACL_DST);
+
+    ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, dst);
+
+    CLAuxTensorHandler lhs_reshaped(offset_int_vec(LhsReshape), _tmp_a, tensors, true);
+    CLAuxTensorHandler rhs_reshaped(offset_int_vec(RhsReshape), _tmp_b, tensors, true);
+
+    // Prepare the consts if needed
+    prepare(tensors);
+
+    // Run matrix multiply kernel
+    switch(_gemm_kernel_type)
+    {
+        case CLGEMMKernelType::NATIVE_V1:
+        {
+            CLScheduler::get().enqueue_op(*_mm_kernel, tensors, true);
+            break;
+        }
+        case CLGEMMKernelType::RESHAPED_V1:
+        case CLGEMMKernelType::RESHAPED:
+        {
+            // Run interleave kernel
+            ITensorPack reshape_lhs_pack{ { ACL_SRC, lhs }, { ACL_DST, lhs_reshaped.get() } };
+            CLScheduler::get().enqueue_op(*_reshape_lhs_kernel, reshape_lhs_pack, false);
+
+            if(!_reshape_b_only_on_first_run)
+            {
+                // Run transpose kernel
+                ITensorPack reshape_rhs_pack{ { ACL_SRC, rhs }, { ACL_DST, rhs_reshaped.get() } };
+                CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, false);
+            }
+
+            ITensorPack gemm_reshaped_pack{ { ACL_SRC_0, lhs_reshaped.get() }, { ACL_SRC_1, rhs_reshaped.get() }, { ACL_SRC_2, src2 }, { ACL_DST, dst } };
+            if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED)
+            {
+                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:
+        {
+            if(!_reshape_b_only_on_first_run)
+            {
+                // Run transpose kernel
+                ITensorPack reshape_rhs_pack{ { ACL_SRC, rhs }, { ACL_DST, rhs_reshaped.get() } };
+                CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, false);
+            }
+            // In case of RESHAPED_ONLY_RHS, we need to check the padding requirement
+            // Check if the lhs or dst tensors have padding
+            const unsigned int cross_plane_pad_lhs = lhs->info()->padding().top + lhs->info()->padding().bottom;
+            const unsigned int cross_plane_pad_dst = dst->info()->padding().top + dst->info()->padding().bottom;
+            bool               has_pad_y           = (cross_plane_pad_lhs != 0) || (cross_plane_pad_dst != 0);
+
+            ITensorPack gemm_reshaped_onlyrhs_pack{ { ACL_SRC_0, lhs }, { ACL_SRC_1, rhs_reshaped.get() }, { ACL_SRC_2, src2 }, { ACL_DST, dst } };
+            if(has_pad_y)
+            {
+                CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_fallback_kernel, gemm_reshaped_onlyrhs_pack, true);
+            }
+            else
+            {
+                CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_kernel, gemm_reshaped_onlyrhs_pack, true);
+            }
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("GEMMType not supported");
+        }
+    }
+}
+
+void ClGemm::prepare(ITensorPack &constants)
+{
+    const ITensor *src1    = constants.get_const_tensor(ACL_SRC_1);
+    ICLTensor     *rhs_aux = utils::cast::polymorphic_downcast<ICLTensor *>(constants.get_tensor(offset_int_vec(RhsReshape)));
+
+    // If memory for RHS is persistent and src1 is provided re-transform else assume that RHS is transformed
+    if((_aux_mem[AuxTensorIdx::RhsReshape].lifetime == MemoryLifetime::Persistent) && (src1 != nullptr && rhs_aux != nullptr) && rhs_aux)
+    {
+        CLAuxTensorHandler rhs_reshaped(_tmp_b, *rhs_aux);
+        ARM_COMPUTE_ERROR_ON(rhs_reshaped.get()->cl_buffer().get() == nullptr);
+
+        ITensorPack reshape_rhs_pack{ { ACL_SRC, src1 }, { ACL_DST, rhs_reshaped.get() } };
+        CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, true);
+    }
+}
+
+experimental::MemoryRequirements ClGemm::workspace() const
+{
+    return _aux_mem;
+}
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/runtime/gpu/cl/operators/ClGemm.h b/src/runtime/gpu/cl/operators/ClGemm.h
new file mode 100644
index 0000000000..bd9ca17edf
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClGemm.h
@@ -0,0 +1,136 @@
+/*
+ * Copyright (c) 2016-2021 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_H
+#define ARM_COMPUTE_CL_GEMM_H
+
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTypes.h"
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.h"
+#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h"
+#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h"
+#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h"
+#include "src/core/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h"
+#include "src/core/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h"
+#include "src/runtime/gpu/cl/IClOperator.h"
+
+#include <memory>
+
+namespace arm_compute
+{
+namespace opencl
+{
+/** 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::ClGemmMatrixMultiplyReshapedOnlyRhsKernel (only if RESHAPED_ONLY_RHS is selected by the select_gemm_kernel method())
+ */
+class ClGemm : public IClOperator
+{
+public:
+    /** Constructor */
+    ClGemm();
+    /** Initialise the kernel's inputs and output
+     *
+     * Valid data layouts:
+     * - All
+     *
+     * Valid data type configurations:
+     * |src0         |src1        |src2      |dst            |
+     * |:------------|:-----------|:---------|:--------------|
+     * |F32          |F32         |F32       |F32            |
+     * |F16          |F16         |F16       |F16            |
+     *
+     * @note GEMM: General Matrix Multiply - [alpha * A * B + beta * C].
+     *
+     * @note All tensors must have the same data type.
+     *
+     * @note Whilst the first input tensor can be a vector, the second input tensor must be at least a matrix
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  a               First input tensor  (Matrix or Vector A). Data types supported: F16/F32
+     * @param[in]  b               Second input tensor (Matrix B). Data type supported: same as @p a.
+     * @param[in]  c               Third input tensor  (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a.
+     * @param[out] output          Output tensor. Data type supported: same as @p a
+     * @param[in]  alpha           Weight of the matrix product
+     * @param[in]  beta            Weight of matrix C
+     * @param[in]  gemm_info       (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
+     *                       if the reshape of matrix B should happen only for the first run. GEMMInfo also contains information about the reshaping
+     *                       in case matrix A and matrix B have been already transformed.
+     */
+    void configure(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to ClGemm::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info);
+
+    // Inherited methods overridden:
+    void run(ITensorPack &tensors) override;
+    void prepare(ITensorPack &constants) override;
+    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_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_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);
+
+private:
+    enum AuxTensorIdx
+    {
+        LhsReshape = 0,
+        RhsReshape,
+        Count
+    };
+
+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::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;
+    TensorInfo                                                          _tmp_a;
+    TensorInfo                                                          _tmp_b;
+    bool                                                                _reshape_b_only_on_first_run;
+    CLGEMMKernelType                                                    _gemm_kernel_type;
+
+    experimental::MemoryRequirements _aux_mem{};
+};
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CLGEMM_H */
diff --git a/src/runtime/gpu/cl/utils/ClAuxTensorHandler.h b/src/runtime/gpu/cl/utils/ClAuxTensorHandler.h
new file mode 100644
index 0000000000..ad893acea5
--- /dev/null
+++ b/src/runtime/gpu/cl/utils/ClAuxTensorHandler.h
@@ -0,0 +1,86 @@
+/*
+ * Copyright (c) 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_UTILS_CL_AUX_TENSOR_HANDLER_H
+#define ARM_COMPUTE_CL_UTILS_CL_AUX_TENSOR_HANDLER_H
+
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+
+#include "support/Cast.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+/* Tensor handler to wrap and handle tensor allocations on workspace buffers */
+class CLAuxTensorHandler
+{
+public:
+    CLAuxTensorHandler(int slot_id, TensorInfo &info, ITensorPack &pack, bool pack_inject = false)
+        : _tensor()
+    {
+        _tensor.allocator()->soft_init(info);
+
+        ICLTensor *packed_tensor = utils::cast::polymorphic_downcast<ICLTensor *>(pack.get_tensor(slot_id));
+        if((packed_tensor == nullptr) || (info.total_size() > packed_tensor->info()->total_size()))
+        {
+            _tensor.allocator()->allocate();
+            if(pack_inject)
+            {
+                pack.add_tensor(slot_id, &_tensor);
+            }
+        }
+        else
+        {
+            _tensor.allocator()->import_memory(packed_tensor->cl_buffer());
+        }
+    }
+
+    CLAuxTensorHandler(TensorInfo &info, ICLTensor &tensor)
+        : _tensor()
+    {
+        _tensor.allocator()->soft_init(info);
+        if(info.total_size() <= tensor.info()->total_size())
+        {
+            _tensor.allocator()->import_memory(tensor.cl_buffer());
+        }
+    }
+
+    ICLTensor *get()
+    {
+        return &_tensor;
+    }
+
+    ICLTensor *operator()()
+    {
+        return &_tensor;
+    }
+
+private:
+    CLTensor _tensor{};
+};
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CL_UTILS_CL_AUX_TENSOR_HANDLER_H */
\ No newline at end of file