* 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.
*/
#include <benchmark/benchmark.h>
#include "faker/fake_value.h"
#include "common/share_graph.h"
#include "runtime/dev.h"
#include "stub/gert_runtime_stub.h"
#include "graph/types.h"
#include "register/kernel_registry.h"
#include "exe_graph/runtime/tensor.h"
#include "graph/utils/graph_utils.h"
#include "lowering/graph_converter.h"
#include "faker/global_data_faker.h"
#include "runtime/model_v2_executor.h"
#include "framework/runtime/executor_option/multi_thread_executor_option.h"
#include "core/executor/multi_thread_topological/executor/schedule/config/task_scheduler_config.h"
#include "lowering/model_converter.h"
#include "kernel/memory/caching_mem_allocator.h"
namespace gert {
namespace {
ge::graphStatus DefaultKernelFunc(KernelContext *run_context) {
return ge::GRAPH_SUCCESS;
}
ge::graphStatus CreateOutput(const ge::Node *node, KernelContext *run_context) {
return ge::GRAPH_SUCCESS;
}
ge::graphStatus freeOutput(const ge::Node *node, KernelContext *run_context) {
return ge::GRAPH_SUCCESS;
}
struct FakeKernelRegistry : KernelRegistry {
const KernelFuncs *FindKernelFuncs(const std::string &kernel_type) const override {
static KernelFuncs funcs = {DefaultKernelFunc, CreateOutput, freeOutput};
return &funcs;
}
} kernel_registry;
const std::string ReduceSumStubName = "ReduceSumStubBin";
const char *const TransDataStubName = "TransDataStubBin";
const char *const TransData13StubName = "TransData17StubBin";
const char *const DynamicAtomicStubName = "DynamicAtomicBin";
const char *const DynamicRnnv3StubName = "DynamicRNNV3StubBin";
const char *const AddStubName = "AddStubBin";
const char *const MulStubName = "MulStubBin";
ge::ComputeGraphPtr GenerateLstmpExeGraph() {
auto graph = ShareGraph::LstmpGraph();
graph->TopologicalSorting();
GE_DUMP(graph, "LstmpST_ComputeGraph");
GeModelBuilder builder(graph);
auto ge_root_model =
builder
.AddTaskDef("TransData",
AiCoreTaskDefFaker(TransDataStubName).AtomicStubNum(DynamicAtomicStubName).WithHandle())
.AddTaskDef("DynamicRNNV3", AiCoreTaskDefFaker(DynamicRnnv3StubName))
.BuildGeRootModel();
auto exe_graph = ModelConverter().ConvertGeModelToExecuteGraph(ge_root_model);
GE_ASSERT_NOTNULL(exe_graph);
GE_DUMP(exe_graph, "LstmpST_ExecuteGraph1");
return exe_graph;
}
}
static void ExecutorRunForLstmpExeGraph(benchmark::State &state) {
auto exe_graph = GenerateLstmpExeGraph();
ge::GraphUtils::DumpGEGraphToOnnx(*exe_graph, "E2ELstmpUT");
KernelRegistry::ReplaceKernelRegistry(std::make_shared<FakeKernelRegistry>());
auto model_executor = ModelV2Executor::Create(exe_graph);
model_executor->Load();
auto outputs = FakeTensors({2}, 3);
auto i0 =
FakeValue<Tensor>(Tensor{{{256}, {256}}, {ge::FORMAT_ND, ge::FORMAT_ND, {}}, kOnDeviceHbm, ge::DT_FLOAT16, 0});
auto i1 =
FakeValue<Tensor>(Tensor{{{256}, {256}}, {ge::FORMAT_ND, ge::FORMAT_ND, {}}, kOnDeviceHbm, ge::DT_FLOAT16, 0});
auto i2 =
FakeValue<Tensor>(Tensor{{{256}, {256}}, {ge::FORMAT_ND, ge::FORMAT_ND, {}}, kOnDeviceHbm, ge::DT_FLOAT16, 0});
auto i3 = FakeValue<uint64_t>(0);
model_executor->Execute({i3.value}, std::vector<Tensor *>({i0.holder.get(), i1.holder.get(), i2.holder.get()}).data(),
3, reinterpret_cast<Tensor **>(outputs.GetAddrList()), outputs.size());
for (auto _ : state) {
model_executor->Execute({i3.value},
std::vector<Tensor *>({i0.holder.get(), i1.holder.get(), i2.holder.get()}).data(), 3,
reinterpret_cast<Tensor **>(outputs.GetAddrList()), outputs.size());
}
model_executor->UnLoad();
KernelRegistry::ReplaceKernelRegistry(nullptr);
}
BENCHMARK(ExecutorRunForLstmpExeGraph);
static void ExecutorWithKernelRunForLstmpExeGraph(benchmark::State &state) {
auto graph = ShareGraph::LstmpGraph();
graph->TopologicalSorting();
GeModelBuilder builder(graph);
auto ge_root_model =
builder
.AddTaskDef("TransData",
AiCoreTaskDefFaker(TransDataStubName).AtomicStubNum(DynamicAtomicStubName).WithHandle())
.AddTaskDef("DynamicRNNV3", AiCoreTaskDefFaker(DynamicRnnv3StubName))
.BuildGeRootModel();
bg::ValueHolder::PopGraphFrame();
auto exe_graph = ModelConverter().ConvertGeModelToExecuteGraph(ge_root_model);
GertRuntimeStub runtime_stub;
runtime_stub.GetKernelStub().StubTiling();
auto model_executor = ModelV2Executor::Create(exe_graph);
auto allocator = memory::CachingMemAllocator::GetAllocator();
auto mem_block = allocator->Malloc(2048 * 2);
memset_s(const_cast<void *>(mem_block->GetAddr()), mem_block->GetSize(), 0, 2048 * 2);
model_executor->Load();
auto outputs = FakeTensors({2048}, 3, const_cast<void *>(mem_block->GetAddr()));
auto i0 = FakeValue<Tensor>(Tensor{{{2048}, {2048}},
{ge::FORMAT_ND, ge::FORMAT_ND, {}},
kOnHost,
ge::DT_FLOAT16,
const_cast<void *>(mem_block->GetAddr())});
auto i1 = FakeValue<Tensor>(Tensor{{{2048}, {2048}},
{ge::FORMAT_ND, ge::FORMAT_ND, {}},
kOnHost,
ge::DT_FLOAT16,
const_cast<void *>(mem_block->GetAddr())});
auto i2 = FakeValue<Tensor>(Tensor{{{2048}, {2048}},
{ge::FORMAT_ND, ge::FORMAT_ND, {}},
kOnHost,
ge::DT_FLOAT16,
const_cast<void *>(mem_block->GetAddr())});
rtStream_t stream;
rtStreamCreate(&stream, static_cast<int32_t>(RT_STREAM_PRIORITY_DEFAULT));
auto i3 = FakeValue<uint64_t>(reinterpret_cast<uint64_t>(stream));
model_executor->Execute({i3.value}, std::vector<Tensor *>({i0.holder.get(), i1.holder.get(), i2.holder.get()}).data(),
3, reinterpret_cast<Tensor **>(outputs.GetAddrList()), outputs.size());
for (auto _ : state) {
model_executor->Execute({i3.value},
std::vector<Tensor *>({i0.holder.get(), i1.holder.get(), i2.holder.get()}).data(), 3,
reinterpret_cast<Tensor **>(outputs.GetAddrList()), outputs.size());
}
model_executor->UnLoad();
rtStreamDestroy(stream);
mem_block->Free();
}
BENCHMARK(ExecutorWithKernelRunForLstmpExeGraph);
static void ParallelExecutorWithKernelRunForLstmpExeGraph(benchmark::State &state) {
auto graph = ShareGraph::LstmpGraph();
graph->TopologicalSorting();
GeModelBuilder builder(graph);
auto ge_root_model =
builder
.AddTaskDef("TransData",
AiCoreTaskDefFaker(TransDataStubName).AtomicStubNum(DynamicAtomicStubName).WithHandle())
.AddTaskDef("DynamicRNNV3", AiCoreTaskDefFaker(DynamicRnnv3StubName))
.BuildGeRootModel();
bg::ValueHolder::PopGraphFrame();
auto exe_graph = ModelConverter().ConvertGeModelToExecuteGraph(ge_root_model);
GertRuntimeStub runtime_stub;
runtime_stub.GetKernelStub().StubTiling();
auto option = MultiThreadExecutorOption(kLeastThreadNumber);
auto model_executor = ModelV2Executor::Create(exe_graph, option);
auto allocator = memory::CachingMemAllocator::GetAllocator();
auto mem_block = allocator->Malloc(2048 * 2);
memset_s(const_cast<void *>(mem_block->GetAddr()), mem_block->GetSize(), 0, 2048 * 2);
model_executor->Load();
auto outputs = FakeTensors({2048}, 3, const_cast<void *>(mem_block->GetAddr()));
auto i0 = FakeValue<Tensor>(Tensor{{{2048}, {2048}},
{ge::FORMAT_ND, ge::FORMAT_ND, {}},
kOnHost,
ge::DT_FLOAT16,
const_cast<void *>(mem_block->GetAddr())});
auto i1 = FakeValue<Tensor>(Tensor{{{2048}, {2048}},
{ge::FORMAT_ND, ge::FORMAT_ND, {}},
kOnHost,
ge::DT_FLOAT16,
const_cast<void *>(mem_block->GetAddr())});
auto i2 = FakeValue<Tensor>(Tensor{{{2048}, {2048}},
{ge::FORMAT_ND, ge::FORMAT_ND, {}},
kOnHost,
ge::DT_FLOAT16,
const_cast<void *>(mem_block->GetAddr())});
rtStream_t stream;
rtStreamCreate(&stream, static_cast<int32_t>(RT_STREAM_PRIORITY_DEFAULT));
auto i3 = FakeValue<uint64_t>(reinterpret_cast<uint64_t>(stream));
model_executor->Execute({i3.value},
std::vector<Tensor *>({i0.holder.get(), i1.holder.get(), i2.holder.get()}).data(),
3, reinterpret_cast<Tensor **>(outputs.GetAddrList()), outputs.size());
for (auto _ : state) {
model_executor->Execute({i3.value},
std::vector<Tensor *>({i0.holder.get(), i1.holder.get(), i2.holder.get()}).data(), 3,
reinterpret_cast<Tensor **>(outputs.GetAddrList()), outputs.size());
}
model_executor->UnLoad();
rtStreamDestroy(stream);
mem_block->Free();
}
BENCHMARK(ParallelExecutorWithKernelRunForLstmpExeGraph);
static void ExecutorWithKernelRunForLstmpExeGraph(benchmark::State &state) {
auto exe_graph = GenerateLstmpExeGraph();
ge::GraphUtils::DumpGEGraphToOnnx(*exe_graph, "E2ELstmpWithKernel");
auto model_executor = ModelV2Executor::Create(exe_graph);
model_executor->Load();
void* device_block = (void*)0x01;
auto outputs = FakeTensors({2}, 3, device_block);
auto i0 = FakeValue<Tensor>(Tensor{{{2048}, {2048}}, {ge::FORMAT_ND, ge::FORMAT_ND, {}}, kOnDeviceHbm, ge::DT_FLOAT16,
device_block}); auto i1 = FakeValue<Tensor>(Tensor{{{2048}, {2048}}, {ge::FORMAT_ND, ge::FORMAT_ND, {}}, kOnDeviceHbm,
ge::DT_FLOAT16, device_block}); auto i2 = FakeValue<Tensor>(Tensor{{{2048}, {2048}}, {ge::FORMAT_ND, ge::FORMAT_ND, {}},
kOnDeviceHbm, ge::DT_FLOAT16, device_block}); auto i3 = FakeValue<uint64_t>(1);
model_executor->Execute({i3.value},
std::vector<Tensor *>({i0.holder.get(), i1.holder.get(), i2.holder.get()}).data(),
3, reinterpret_cast<Tensor **>(outputs.GetAddrList()), outputs.size()); for (auto _ : state) {
model_executor->Execute({i3.value},
std::vector<Tensor *>({i0.holder.get(), i1.holder.get(), i2.holder.get()}).data(),
3, reinterpret_cast<Tensor **>(outputs.GetAddrList()), outputs.size());
}
model_executor->UnLoad();
}
BENCHMARK(ExecutorWithKernelRunForLstmpExeGraph);
*/
}
BENCHMARK_MAIN();