root
/
MNN
mirror of https://github.com/alibaba/MNN.git
1
0
Fork 0
Code Issues Actions 1 Packages Projects Releases Wiki Activity

refactor cpuids setting from BackendConfig to HintMode

This commit is contained in:
Jules 2025-07-18 09:42:59 +00:00
parent 8907c96b4e
commit c018eacc00
14 changed files with 96 additions and 70 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
docs/inference/session.md
View File

@ -185,13 +185,10 @@ struct BackendConfig {
void* sharedContext = nullptr;
size_t flags; // Valid for CPU Backend
};
/** user specified cpu cores */
std::vector<int> cpuIds;
};
```
`memory`、`power`、`precision`分别为内存、功耗和精度偏好。支持这些选项的后端会在执行时做出相应调整；若不支持，则忽略选项。`cpuIds`允许用户指定一组CPU核心用于计算，但这些指定会被严格的校验合法性，当线程数量与合法的指定核心数量相同时得到最佳性能。
`memory`、`power`、`precision`分别为内存、功耗和精度偏好。支持这些选项的后端会在执行时做出相应调整；若不支持，则忽略选项。
示例：
后端 **OpenCL**

8
express/Executor.cpp
View File

@ -231,6 +231,14 @@ void Executor::RuntimeManager::setHint(Interpreter::HintMode mode, int value) {
iter.second->setRuntimeHint(mInside->mContent->modes.runtimeHint);
}
}
void Executor::RuntimeManager::setHint(Interpreter::HintMode mode, int* value, size_t size) {
mInside->mContent->modes.setHint(mode, value, size);
auto current = ExecutorScope::Current();
auto rt = current->getRuntime();
for (auto& iter : rt.first) {
iter.second->setRuntimeHint(mInside->mContent->modes.runtimeHint);
}
}
void Executor::RuntimeManager::setExternalPath(std::string path, int type) {
mInside->mContent->modes.setExternalPath(path, type);
}

11
include/MNN/Interpreter.hpp
View File

@ -245,7 +245,10 @@ public:
USE_CACHED_MMAP = 12,
// Multi-Thread Load module, default is 0 (don't use other Thread)
INIT_THREAD_NUMBER = 13
INIT_THREAD_NUMBER = 13,
// CPU core ids
CPU_CORE_IDS = 14,
};
enum ExternalPathType {
@ -280,10 +283,12 @@ public:
/**
* @brief The API shoud be called before create session.
* @param mode Hint type
* @param hint Hint type
* @param value Hint value
* @param size Hint value size(when use a ptr)
*/
void setSessionHint(HintMode mode, int value);
void setSessionHint(HintMode hint, int value);
void setSessionHint(HintMode hint, int* value, size_t size);
public:
/**
* @brief create runtimeInfo separately with schedule config.

4
include/MNN/MNNForwardType.h
View File

@ -10,7 +10,6 @@
#define MNNForwardType_h
#include <stdint.h>
#include <stddef.h>
#include <vector>
typedef enum {
MNN_FORWARD_CPU = 0,
@ -93,9 +92,6 @@ struct BackendConfig {
void* sharedContext = nullptr;
size_t flags; // Valid for CPU Backend
};
/** user specified cpu cores */
std::vector<int> cpuIds;
};
/** acquire runtime status by Runtime::getCurrentStatus with following keys,

1
include/MNN/expr/Executor.hpp
View File

@ -126,6 +126,7 @@ public:
friend class Executor;
void setMode(Interpreter::SessionMode mode);
void setHint(Interpreter::HintMode mode, int value);
void setHint(Interpreter::HintMode mode, int* value, size_t size);
void setHintPtr(Interpreter::HintMode mode, void* value);
bool getInfo(Interpreter::SessionInfoCode code, void* ptr);
BackendConfig* getBnConfig();

19
source/backend/cpu/CPUBackend.cpp
View File

@ -113,7 +113,7 @@ void CPURuntime::_bindCPUCore() const {
#endif
}
void CPURuntime::_resetThreadPool() {
void CPURuntime::_resetThreadPool() const{
mThreadNumber = std::max(1, mThreadNumber);
mThreadNumber = std::min(mThreadNumber, MAX_THREAD_NUMBER);
#ifdef MNN_USE_THREAD_POOL
@ -136,7 +136,7 @@ void CPURuntime::_resetThreadPool() {
// Reset tid to rebind cpu if necessary
mCurrentTID = 0;
}
void CPURuntime::_validateCpuIds() {
void CPURuntime::_validateCpuIds() const{
bool valid = true;
do {
@ -200,8 +200,7 @@ void CPURuntime::_validateCpuIds() {
case BackendConfig::Power_Low:
mCpuIds = cpuInfo->groups[0].ids;
break;
case BackendConfig::Power_High:
{
case BackendConfig::Power_High: {
int selectCPUSize = 0;
int groupIndex = cpuInfo->groups.size() - 1;
while (selectCPUSize < mThreadNumber && groupIndex >= 0) {
@ -220,7 +219,6 @@ void CPURuntime::_validateCpuIds() {
void CPURuntime::onReset(int numberThread, const BackendConfig* config, bool full) {
if (config != nullptr) {
mPower = config->power;
mCpuIds = config->cpuIds;
if (full) {
mPrecision = config->precision;
mMemory = config->memory;
@ -228,6 +226,7 @@ void CPURuntime::onReset(int numberThread, const BackendConfig* config, bool ful
}
}
mThreadNumber = numberThread;
mCpuIds = hint().cpuIds;
_validateCpuIds();
mCpuMask = MNNGetCPUMask(mCpuIds);
_resetThreadPool();
@ -250,11 +249,7 @@ CPURuntime::CPURuntime(const Backend::Info& info) {
mPower = info.user->power;
mMemory = info.user->memory;
mFlags = info.user->flags;
mCpuIds = info.user->cpuIds;
}
_validateCpuIds();
mCpuMask = MNNGetCPUMask(mCpuIds);
_resetThreadPool();
#ifdef LOG_VERBOSE
MNN_PRINT("create CPURuntime:%p\n", this);
#endif
@ -287,6 +282,12 @@ SingleBufferWithAllocator* CPURuntime::buffer(int index) const {
}
Backend* CPURuntime::onCreate(const BackendConfig* config, Backend* origin) const {
{
mCpuIds = hint().cpuIds;
_validateCpuIds();
mCpuMask = MNNGetCPUMask(mCpuIds);
_resetThreadPool();
}
if (hint().midMemoryPath.size() > 0) {
if (mDynamicMmap.empty()) {
// Only support set featuremap dir once

12
source/backend/cpu/CPUBackend.hpp
View File

@ -54,16 +54,16 @@ public:
private:
void _bindCPUCore() const;
void _resetThreadPool();
void _validateCpuIds();
void _resetThreadPool() const;
void _validateCpuIds() const;
mutable std::shared_ptr<EagerBufferAllocator> mStaticAllocator;
int mThreadNumber;
std::vector<int> mCpuIds;
unsigned long mCpuMask;
mutable int mThreadNumber;
mutable std::vector<int> mCpuIds;
mutable unsigned long mCpuMask;
#ifdef MNN_USE_THREAD_POOL
mutable int mTaskIndex = -1;
mutable int mThreadOpen = 0;
ThreadPool* mThreadPool = nullptr;
mutable ThreadPool* mThreadPool = nullptr;
#endif
BackendConfig::MemoryMode mMemory;
BackendConfig::PowerMode mPower;

3
source/core/Backend.hpp
View File

@ -58,6 +58,9 @@ struct RuntimeHint {
// op encoder number for once commit
int encorderNumForCommit = 10;
int initThreadNumber = 0;
// cpu core ids
std::vector<int> cpuIds;
};
/** abstract backend */
class Backend : public NonCopyable {

8
source/core/Interpreter.cpp
View File

@ -140,8 +140,12 @@ Interpreter* Interpreter::createFromBufferInternal(Content* net, bool enforceAut
return new Interpreter(net);
}
void Interpreter::setSessionHint(HintMode mode, int hint) {
mNet->modes.setHint(mode, hint);
void Interpreter::setSessionHint(HintMode hint, int value) {
mNet->modes.setHint(hint, value);
}
void Interpreter::setSessionHint(HintMode hint, int* value, size_t size) {
mNet->modes.setHint(hint, value, size);
}
void Interpreter::setSessionMode(SessionMode mode) {

65
source/core/Session.cpp
View File

@ -68,46 +68,55 @@ void Session::ModeGroup::setMode(Interpreter::SessionMode mode) {
codegenMode = mode;
}
}
void Session::ModeGroup::setHint(Interpreter::HintMode mode, int hint) {
switch (mode) {
case Interpreter::MAX_TUNING_NUMBER:
maxTuningNumber = hint;
void Session::ModeGroup::setHint(Interpreter::HintMode hint, int value) {
switch (hint) {
case Interpreter::HintMode::MAX_TUNING_NUMBER:
maxTuningNumber = value;
break;
case Interpreter::MEM_ALLOCATOR_TYPE:
runtimeHint.memoryAllocatorType = hint;
case Interpreter::HintMode::MEM_ALLOCATOR_TYPE:
runtimeHint.memoryAllocatorType = value;
break;
case Interpreter::WINOGRAD_MEMORY_LEVEL:
runtimeHint.winogradMemoryUsed = hint;
case Interpreter::HintMode::WINOGRAD_MEMORY_LEVEL:
runtimeHint.winogradMemoryUsed = value;
break;
case Interpreter::CPU_LITTLECORE_DECREASE_RATE:
runtimeHint.cpuDecreaseRate = hint;
case Interpreter::HintMode::CPU_LITTLECORE_DECREASE_RATE:
runtimeHint.cpuDecreaseRate = value;
break;
case Interpreter::GEOMETRY_COMPUTE_MASK:
geometryMask = hint;
case Interpreter::HintMode::GEOMETRY_COMPUTE_MASK:
geometryMask = value;
break;
case Interpreter::STRICT_CHECK_MODEL:
checkNetBuffer = hint > 0;
case Interpreter::HintMode::STRICT_CHECK_MODEL:
checkNetBuffer = value > 0;
break;
case Interpreter::DYNAMIC_QUANT_OPTIONS:
runtimeHint.dynamicQuantOption = hint;
case Interpreter::HintMode::DYNAMIC_QUANT_OPTIONS:
runtimeHint.dynamicQuantOption = value;
break;
case Interpreter::QKV_QUANT_OPTIONS:
runtimeHint.qkvQuantOption = hint;
case Interpreter::HintMode::QKV_QUANT_OPTIONS:
runtimeHint.qkvQuantOption = value;
break;
case Interpreter::KVCACHE_SIZE_LIMIT:
runtimeHint.kvcacheSizeLimit = hint;
case Interpreter::HintMode::KVCACHE_SIZE_LIMIT:
runtimeHint.kvcacheSizeLimit = value;
break;
case Interpreter::OP_ENCODER_NUMBER_FOR_COMMIT:
runtimeHint.encorderNumForCommit = hint;
case Interpreter::HintMode::OP_ENCODER_NUMBER_FOR_COMMIT:
runtimeHint.encorderNumForCommit = value;
break;
case Interpreter::MMAP_FILE_SIZE:
runtimeHint.mmapFileSize = hint;
case Interpreter::HintMode::MMAP_FILE_SIZE:
runtimeHint.mmapFileSize = value;
break;
case Interpreter::USE_CACHED_MMAP:
runtimeHint.useCachedMmap = hint;
case Interpreter::HintMode::USE_CACHED_MMAP:
runtimeHint.useCachedMmap = value;
break;
case Interpreter::INIT_THREAD_NUMBER:
runtimeHint.initThreadNumber = hint;
case Interpreter::HintMode::INIT_THREAD_NUMBER:
runtimeHint.initThreadNumber = value;
break;
default:
break;
}
}
void Session::ModeGroup::setHint(Interpreter::HintMode hint, int* value, size_t size) {
switch (hint) {
case Interpreter::HintMode::CPU_CORE_IDS:
runtimeHint.cpuIds = std::vector<int>(value, value + size);
break;
default:
break;

4
source/core/Session.hpp
View File

@ -37,7 +37,9 @@ public:
int geometryMask = 0xFFFF;
bool checkNetBuffer = true;
RuntimeHint runtimeHint;
void setHint(Interpreter::HintMode hint, int magic);
void setHint(Interpreter::HintMode hint, int value);
void setHint(Interpreter::HintMode hint, int* value, size_t size);
void setHintPtr(Interpreter::HintMode hint, int value);
void setMode(Interpreter::SessionMode mode);
void setExternalPath(std::string path, int type);
};

2
tools/cpp/MNNV2Basic.cpp
View File

@ -282,6 +282,7 @@ static int test_main(int argc, const char* argv[]) {
if (runMask & 32) {
net->setSessionHint(Interpreter::WINOGRAD_MEMORY_LEVEL, 0);
}
net->setSessionHint(Interpreter::HintMode::CPU_CORE_IDS, cpuIds.data(), cpuIds.size());
// create session
MNN::ScheduleConfig config;
@ -295,7 +296,6 @@ static int test_main(int argc, const char* argv[]) {
// backendConfig.power = BackendConfig::Power_High;
backendConfig.precision = static_cast<MNN::BackendConfig::PrecisionMode>(precision);
backendConfig.memory = static_cast<MNN::BackendConfig::MemoryMode>(memory);
backendConfig.cpuIds = cpuIds;
config.backendConfig = &backendConfig;
MNN::Session* session = NULL;
MNN::Tensor* inputTensor = nullptr;

2
tools/cpp/ModuleBasic.cpp
View File

@ -264,7 +264,6 @@ int main(int argc, char *argv[]) {
backendConfig.power = (BackendConfig::PowerMode)power;
backendConfig.precision = static_cast<MNN::BackendConfig::PrecisionMode>(precision);
backendConfig.memory = static_cast<MNN::BackendConfig::MemoryMode>(memory);
backendConfig.cpuIds = cpuIds;
config.backendConfig = &backendConfig;
MNN::Express::Module::Config mConfig;
@ -275,6 +274,7 @@ int main(int argc, char *argv[]) {
std::shared_ptr<Executor::RuntimeManager> rtmgr(Executor::RuntimeManager::createRuntimeManager(config));
rtmgr->setCache(cacheFileName);
rtmgr->setHint(MNN::Interpreter::INIT_THREAD_NUMBER, 4);
rtmgr->setHint(MNN::Interpreter::HintMode::CPU_CORE_IDS, cpuIds.data(), cpuIds.size());
if (cpuDecreaseRate > 0 && cpuDecreaseRate <= 100) {
rtmgr->setHint(Interpreter::CPU_LITTLECORE_DECREASE_RATE, cpuDecreaseRate);

2
tools/cpp/timeProfile.cpp
View File

@ -119,6 +119,7 @@ int main(int argc, const char* argv[]) {
}
revertor.reset();
net->setSessionMode(Interpreter::Session_Debug);
net->setSessionHint(Interpreter::HintMode::CPU_CORE_IDS, cpuIds.data(), cpuIds.size());
// create session
MNN::ScheduleConfig config;
@ -126,7 +127,6 @@ int main(int argc, const char* argv[]) {
config.numThread = threadNumber;
BackendConfig backendConfig;
backendConfig.precision = precision;
backendConfig.cpuIds = cpuIds;
config.backendConfig = &backendConfig;
MNN::Session* session = NULL;
session = net->createSession(config);