* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#ifndef JIT_EXECUTOR_H
#define JIT_EXECUTOR_H
#include <vector>
#include <queue>
#include <mutex>
#include <map>
#include <set>
#include "acl/acl_rt.h"
#include "exe_graph/runtime/runtime_tensor.h"
#include "ge/ge_api_types.h"
#include "exe_points/execution_order.h"
#include "cache/compiled_model_cache.h"
#include "compile_context.h"
#include "graph/utils/tensor_adapter.h"
#include "graph/utils/type_utils.h"
namespace ge {
struct UserGraphExecution {
UserGraphExecution(uint32_t graph_id, const std::vector<gert::Tensor> &inputs,
const RunAsyncCallbackV2 &callback_func, uint64_t session_id_param)
: user_graph_id(graph_id), callback(callback_func), session_id(session_id_param), external_rt_inputs(&inputs) {
inputs_memblocks.resize(inputs.size());
}
UserGraphExecution(uint32_t graph_id, std::vector<gert::Tensor> &&inputs, const RunAsyncCallbackV2 &callback_func,
uint64_t session_id_param)
: user_graph_id(graph_id),
callback(callback_func),
session_id(session_id_param),
input_tensors_holder(std::move(inputs)),
external_rt_inputs(&input_tensors_holder) {
inputs_memblocks.resize(input_tensors_holder.size());
}
~UserGraphExecution() = default;
uint32_t user_graph_id;
std::vector<MemBlock *> inputs_memblocks;
RunAsyncCallbackV2 callback{nullptr};
void *stream{nullptr};
uint64_t session_id;
std::vector<gert::Tensor> input_tensors_holder;
const std::vector<gert::Tensor> *external_rt_inputs{nullptr};
std::vector<gert::Tensor> *rt_outputs{nullptr};
std::map<AscendString, AscendString> load_options;
};
using UserGraphExecutionQueue = std::queue<std::unique_ptr<UserGraphExecution>>;
template <typename K, typename V>
std::vector<std::pair<K, V>> SortMapByValue(const std::map<K, V> &input_map, bool is_ascend = true) {
std::vector<std::pair<K, V>> target_vec(input_map.begin(), input_map.end());
std::sort(target_vec.begin(), target_vec.end(), [is_ascend](const std::pair<K, V> &a, const std::pair<K, V> &b) {
return is_ascend ? (a.second < b.second) : (a.second > b.second);
});
return target_vec;
}
class JitExecutor {
public:
static std::unique_ptr<JitExecutor> Create(
GraphManager &graph_manager, UserGraphExecutionQueue &task_queue, ExecutionOrder &order,
CompileContext &compile_context, CompiledModelCache &cmc, std::mutex &mutex,
const std::map<MemoryType, std::pair<const void *, size_t>> *fixed_feature_memory_settings = nullptr);
Status RunWithCallback(UserGraphExecution &&task);
Status Finalize();
bool IsUserGraphNeedRebuild();
Status CompileGraph(UserGraphExecution &task, uint64_t session_id);
Status LoadGraph(UserGraphExecution &task);
Status Execute(UserGraphExecution &&task);
private:
struct GuardedExecutionInfo {
GuardedExecutionPoint *gep{nullptr};
uint32_t instance_id{0U};
};
struct DataNodeInfo {
OpDescPtr op_desc;
int32_t input_index{-1};
};
JitExecutor(GraphManager &graph_manager, UserGraphExecutionQueue &task_queue, ExecutionOrder &order,
CompileContext &compile_context, CompiledModelCache &cmc, std::mutex &mutex,
const std::map<MemoryType, std::pair<const void *, size_t>> *fixed_feature_memory_settings = nullptr);
Status CompileAndLoad(const std::vector<gert::Tensor> &inputs, GuardedExecutionPoint *gep, uint32_t &instance_id,
const aclrtStream stream, const std::map<AscendString, AscendString> &load_options,
uint64_t session_id);
Status ApplyFixedFeatureMemory(uint32_t instance_id) const;
Status Compile(const std::vector<ge::Tensor> &inputs, GuardedExecutionPoint *gep, uint64_t session_id);
Status ProcessAndExecuteGraphAsync(UserGraphExecution &task, aclrtStream const stream,
const std::vector<gert::Tensor> &inputs, std::vector<gert::Tensor> &outputs,
ExecutionPoint *ep, bool need_malloc_output = false);
Status TryExecuteWithoutProcess(UserGraphExecution &task);
Status MallocOutputsForStatic(uint32_t guarded_ep_instance_id, const GuardedExecutionPoint *gep,
std::vector<gert::Tensor> &outputs);
Status GetOrCompileGuardedExecutionPoint(UserGraphExecution &task, const std::vector<gert::Tensor> &compile_inputs,
ExecutionPoint *ep, rtStream_t stream, GuardedExecutionInfo &execution_info);
std::vector<DataNodeInfo> GetOrCreateDataNodeInfos(const ComputeGraphPtr &graph);
void MarkHostTensorOnDataNodes(const std::vector<gert::Tensor> &inputs, const ComputeGraphPtr &graph);
GraphManager &graph_manager_;
UserGraphExecutionQueue &task_queue_;
ExecutionOrder &order_;
CompileContext &compile_context_;
CompiledModelCache &cmc_;
std::mutex &mutex_;
const std::map<MemoryType, std::pair<const void *, size_t>> *fixed_feature_memory_settings_{nullptr};
std::map<const GuardedExecutionPoint *, uint32_t> geps_to_inner_ge_graph_id_;
aclrtStream stream_{nullptr};
std::shared_ptr<ge::Allocator> device_allocator_;
int32_t device_id_{-1};
std::vector<uint32_t> compiled_ge_graph_id_;
std::shared_ptr<ge::Allocator> external_allocator_{nullptr};
struct GuardedExecutionCacheEntry {
bool valid{false};
uint32_t compiled_graph_id{0U};
uint32_t instance_id{0U};
};
std::map<const ExecutionPoint *, GuardedExecutionCacheEntry> guarded_execution_cache_;
std::mutex guarded_execution_cache_mutex_;
std::map<ComputeGraph *, std::set<size_t>> cond_input_data_cache_;
std::map<ComputeGraph *, std::vector<DataNodeInfo>> data_node_cache_;
};
}
#endif
