[GPU] Move tuning cache loading to kernel selector (openvinotoolkit#15112)

* [GPU] Move tuning cache loading to kernel selector. Remove tuning modes

* [GPU] Removed kernel runner

Author: Vladimir Paramuzov

src/plugins/intel_gpu/include/intel_gpu/graph/program.hpp

@@ -17,10 +17,6 @@
 #include <utility>
 #include <set>
 
-namespace kernel_selector {
-class TuningCache;
-}  // namespace kernel_selector
-
 namespace cldnn {
 
 struct topology;
@@ -248,9 +244,6 @@ struct program {
     kernel::ptr get_kernel(kernel_id id);
     kernels_cache& get_kernels_cache() const;
 
-    void load_tuning_cache();
-    std::shared_ptr<kernel_selector::TuningCache> get_tuning_cache() const { return tuning_cache; }
-
     // returns {-1, -1} if it failed to estimate by allocating given batch size
     std::pair<int64_t/*const alloc*/, int64_t/*general alloc*/> get_estimated_device_mem_usage();
 
@@ -270,7 +263,6 @@ struct program {
     std::vector<program_node*> outputs;
     nodes_ordering processing_order;
     std::unique_ptr<pass_manager> pm;
-    std::shared_ptr<kernel_selector::TuningCache> tuning_cache;
     bool is_body_program;
     int8_t is_subgroup_local_block_io_supported;
 

src/plugins/intel_gpu/include/intel_gpu/runtime/internal_properties.hpp

@@ -48,41 +48,6 @@ static constexpr Property<bool, PropertyMutability::RW> partial_build_program{"G
 static constexpr Property<bool, PropertyMutability::RW> allow_new_shape_infer{"GPU_ALLOW_NEW_SHAPE_INFER"};
 static constexpr Property<std::string, PropertyMutability::RW> dump_graphs{"GPU_DUMP_GRAPHS"};
 static constexpr Property<std::vector<std::string>, PropertyMutability::RW> custom_outputs{"GPU_CUSTOM_OUTPUTS"};
-
-/// @brief Tuning mode.
-enum class TuningMode {
-    /// @brief Tuning is disabled.
-    tuning_disabled,
-
-    /// @brief Tuning using the cached data (no on-line tuning for non-existing data).
-    tuning_use_cache,
-
-    /// @brief Tuning using the cached data if exist, tune and update cache otherwise.
-    tuning_tune_and_cache,
-
-    /// @brief Tuning using the cached data and update tasks.
-    /// @details Performs updating tasks like removal of invalid caches, promoting to new format, etc.
-    /// No tuning for non-existing data.
-    tuning_use_and_update,
-
-    /// @brief Retune the cache data even if it exists.
-    tuning_retune_and_cache
-};
-
-struct TuningConfig {
-    TuningMode mode;
-    std::string cache_file_path;
-
-    TuningConfig() : mode(TuningMode::tuning_disabled), cache_file_path("") {}
-};
-
-inline std::ostream& operator<<(std::ostream& os, const TuningConfig& val) {
-    os << val.cache_file_path;
-    return os;
-}
-
-static constexpr Property<TuningConfig, PropertyMutability::RW> tuning_config{"GPU_TUNING_CONFIG"};
-
 static constexpr Property<ImplForcingMap, PropertyMutability::RW> force_implementations{"GPU_FORCE_IMPLEMENTATIONS"};
 static constexpr Property<std::string, PropertyMutability::RW> config_file{"CONFIG_FILE"};
 static constexpr Property<bool, PropertyMutability::RW> enable_lp_transformations{"LP_TRANSFORMS_MODE"};

src/plugins/intel_gpu/src/graph/impls/ocl/arg_max_min.cpp

@@ -9,7 +9,6 @@
 #include "kernel_selector_helper.h"
 #include "arg_max_min/arg_max_min_kernel_selector.h"
 #include "arg_max_min/arg_max_min_kernel_base.h"
-#include "kernel_runner.h"
 
 namespace cldnn {
 namespace ocl {

src/plugins/intel_gpu/src/graph/impls/ocl/binary_convolution.cpp

@@ -8,7 +8,6 @@
 #include "impls/implementation_map.hpp"
 #include "intel_gpu/runtime/error_handler.hpp"
 #include "kernel_selector_helper.h"
-#include "kernel_runner.h"
 #include "kernel_selector/kernels/binary_convolution/binary_convolution_kernel_selector.h"
 #include "kernel_selector/kernels/binary_convolution/binary_convolution_params.h"
 #include <algorithm>
@@ -75,14 +74,6 @@ struct binary_convolution_impl : typed_primitive_impl_ocl<binary_convolution> {
         uint32_t dilation_x = dilation.size() >= 1 ? dilation[dilation.size() - 1] : 1;
         params.dilation = {dilation_x, dilation_y, dilation_z};
 
-        const auto& tuning_config = impl_param.get_program().get_config().get_property(ov::intel_gpu::tuning_config);
-
-        if (tuning_config.mode == ov::intel_gpu::TuningMode::tuning_tune_and_cache ||
-            tuning_config.mode == ov::intel_gpu::TuningMode::tuning_retune_and_cache) {
-            optional_params.tuningParams.runner =
-                std::make_shared<gpu::kernel_runner>(impl_param.get_program().get_engine(), impl_param.get_program().get_id(), true);
-        }
-
         return {params, optional_params};
     }
 };

src/plugins/intel_gpu/src/graph/impls/ocl/convolution.cpp

@@ -8,7 +8,6 @@
 #include "impls/implementation_map.hpp"
 #include "intel_gpu/runtime/error_handler.hpp"
 #include "kernel_selector_helper.h"
-#include "kernel_runner.h"
 #include "convolution/convolution_kernel_selector.h"
 #include "convolution/convolution_params.h"
 #include <algorithm>
@@ -166,14 +165,6 @@ struct convolution_impl : typed_primitive_impl_ocl<convolution> {
 
         auto& kernel_selector = kernel_selector::convolution_kernel_selector::Instance();
 
-        const auto& tuning_config = impl_param.get_program().get_config().get_property(ov::intel_gpu::tuning_config);
-
-        if (tuning_config.mode == ov::intel_gpu::TuningMode::tuning_tune_and_cache ||
-            tuning_config.mode == ov::intel_gpu::TuningMode::tuning_retune_and_cache) {
-            conv_optional_params.tuningParams.runner =
-                std::make_shared<gpu::kernel_runner>(arg.get_program().get_engine(), arg.get_program().get_id(), true, true);
-        }
-
         auto best_kernel = kernel_selector.get_best_kernel(conv_params, conv_optional_params);
 
         return make_unique<convolution_impl>(best_kernel);

src/plugins/intel_gpu/src/graph/impls/ocl/deformable_convolution.cpp

@@ -7,7 +7,6 @@
 #include "impls/implementation_map.hpp"
 #include "intel_gpu/runtime/error_handler.hpp"
 #include "kernel_selector_helper.h"
-#include "kernel_runner.h"
 #include "convolution/convolution_kernel_selector.h"
 #include "convolution/convolution_params.h"
 #include <algorithm>

src/plugins/intel_gpu/src/graph/impls/ocl/fully_connected.cpp

@@ -10,7 +10,6 @@
 #include "fully_connected/fully_connected_params.h"
 
 #include "intel_gpu/runtime/error_handler.hpp"
-#include "kernel_runner.h"
 
 #include "intel_gpu/primitives/reorder.hpp"
 #include "intel_gpu/primitives/input_layout.hpp"
@@ -119,7 +118,6 @@ struct fully_connected_impl : typed_primitive_impl_ocl<fully_connected> {
             params.quantization = kernel_selector::QuantizationType::NONE;
         }
 
-        optional_params.tuningParams.runner = std::make_shared<gpu::kernel_runner>(progam.get_engine(), progam.get_id(), true);
         return {params, optional_params};
     }
 

src/plugins/intel_gpu/src/graph/include/kernel_selector_helper.h

@@ -28,7 +28,6 @@ using namespace cldnn;
 
 namespace cldnn {
 enum class data_types : size_t;
-enum class tuning_mode;
 struct format;
 struct layout;
 struct program;
@@ -65,7 +64,6 @@ using softmax_dim = kernel_selector::SoftmaxDim;
 using mean_subtruct_mode = kernel_selector::MeanSubtractMode;
 using mean_op = kernel_selector::MeanOp;
 using concat_axis = kernel_selector::ConcatAxis;
-using tuning_mode = kernel_selector::TuningMode;
 using sample_type = kernel_selector::ResampleType;
 using coordinate_transformation_mode = kernel_selector::CoordinateTransformationMode;
 using nearest_mode = kernel_selector::NearestMode;
@@ -101,7 +99,6 @@ kernel_selector::data_layout to_data_layout(format f);
 cldnn::format from_data_layout(kernel_selector::data_layout l);
 kernel_selector::weights_layout to_weights_layout(format f, bool is_grouped);
 cldnn::format::type from_weights_layout(kernel_selector::weights_layout l);
-kernel_selector::tuning_mode to_tuning_mode(ov::intel_gpu::TuningMode mode);
 kernel_selector::data_tensor convert_data_tensor(const layout& l, const tensor view_offset = tensor {});
 kernel_selector::weights_tensor convert_weights_tensor(const layout& l, bool is_grouped = false);
 layout from_weights_tensor(const kernel_selector::weights_tensor& t);