Skip to content

Commit

Permalink
[MNN:Sync] Sync Internal 2.8.2
Browse files Browse the repository at this point in the history
  • Loading branch information
wangzhaode committed Feb 29, 2024
1 parent 5607201 commit 970b63f
Show file tree
Hide file tree
Showing 295 changed files with 1,013,758 additions and 3,667 deletions.
5 changes: 4 additions & 1 deletion CMakeLists.txt
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,7 @@ option(MNN_DEBUG_MEMORY "MNN Debug Memory Access" OFF)
option(MNN_DEBUG_TENSOR_SIZE "Enable Tensor Size" OFF)
option(MNN_GPU_TRACE "Enable MNN Gpu Debug" OFF)
option(MNN_SUPPORT_RENDER "Enable MNN Render Ops" OFF)
option(MNN_SUPPORT_TRANSFORMER_FUSE "Enable MNN transformer Fuse Ops" OFF)
option(MNN_PORTABLE_BUILD "Link the static version of third party libraries where possible to improve the portability of built executables" OFF)
option(MNN_SEP_BUILD "Build MNN Backends and expression separately. Only works with MNN_BUILD_SHARED_LIBS=ON" ON)
option(NATIVE_LIBRARY_OUTPUT "Native Library Path" OFF)
Expand Down Expand Up @@ -166,7 +167,9 @@ endif()
if(MNN_SUPPORT_RENDER)
add_definitions(-DMNN_SUPPORT_RENDER)
endif()

if(MNN_SUPPORT_TRANSFORMER_FUSE)
add_definitions(-DMNN_SUPPORT_TRANSFORMER_FUSE)
endif()
# debug options
if(MNN_DEBUG_MEMORY)
add_definitions(-DMNN_DEBUG_MEMORY)
Expand Down
6 changes: 6 additions & 0 deletions docker_release.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,6 @@
# using docker run release
docker start mnn_release
docker exec -i -e TEST_ID=$(pwd | awk -F "/" '{print $(NF-1)}') mnn_release bash <<'EOF'
cd ~/yanxing_zhaode/cise/space/$TEST_ID/source && ./release.sh pymnn
exit
EOF
1 change: 1 addition & 0 deletions docs/compile/cmake.md
Original file line number Diff line number Diff line change
Expand Up @@ -81,4 +81,5 @@ MNN使用CMake构建项目,CMake中的宏定义列表如下:
| MNN_VULKAN_IMAGE | 构建MNN的Vulkan后端时采用Image内存模式,以便支持FP16和部分移动端上GPU的加速,默认为`ON` |
| MNN_LOW_MEMORY | 是否支持低内存模式,支持低内存模式使用权值量化模型并设置`low_memory`则会使用计算时反量化,默认为`OFF` |
| MNN_SUPPORT_RENDER | 是否支持图形渲染相关算子实现,默认为 `OFF` |
| MNN_SUPPORT_TRANSFORMER_FUSE | 是否支持Fuse Transformer相关OP实现,默认为 `OFF` |
| MNN_BUILD_LLM | 是否构建基于MNN的llm库和demo,默认为`OFF` |
4 changes: 2 additions & 2 deletions docs/faq.md
Original file line number Diff line number Diff line change
Expand Up @@ -246,7 +246,7 @@ GPU 后端调用 copy 的时间包含两个部分
- x86 / x64 架构下,无 vnni 指令,量化计算需要先从 int8 转到 int16 ,乘加到 int32 ,本身计算效率不如浮点直接乘加到 fp32 上快。
- x64 + vnni 指令,量化计算有 sdot 指令,明显快于 FP32 ,编译 MNN 时需要开启 MNN_AVX512 以支持这个指令,一般相比 AVX512 的浮点运算快 30%
- ARM v7a / ARMv8 :量化计算采用 int8 乘加到 int16,再双加到 int32 的方式,计算效率略快于浮点(一般 30% 左右提升)。
- ARMv8.2 + 量化计算有 sdot 指令,但同时 FP32 相对之前架构发射数也提升了一倍,编译 MNN 打开 MNN_ARM82 启用 sdot 指令则量化计算更快,否则 FP32 更快,理想情况比 FP32 快1倍以上,比 FP16 快 20%。
- ARMv8.2 架构有 sdot 指令,但同时 FP32 相对之前架构发射数也提升了一倍,也支持了比 FP32 快一倍的 FP16 向量计算指令,MNN 会检查设备架构以开启 sdot / smmla ,理想情况下量化计算性能比 FP32 快1倍以上,比 FP16 快 20%。
## 其他问题
### MNN模型如何加密
Expand All @@ -256,4 +256,4 @@ GPU 后端调用 copy 的时间包含两个部分
2. 执行`schema/generate.sh`重新生成`flatbuffers`头文件;
3. 重新编译`MNN`库文件, `Convert`等所有工具;
4. 使用新的工具重新转换模型;
5. 在端侧使用新模型和新的`MNN`库文件进行部署;
5. 在端侧使用新模型和新的`MNN`库文件进行部署;
7 changes: 5 additions & 2 deletions docs/inference/module.md
Original file line number Diff line number Diff line change
Expand Up @@ -45,9 +45,10 @@ rtmgr->setCache(".cachefile");
// 从模型文件加载并创建新Module
const std::string model_file = "/tmp/mymodule.mnn"; // model file with path
// 输入名,可以为空,为空时 MNN 自动搜索模型中的输入,多输入情况下无法保证顺序,需要通过 getInfo 接口查看
// 输入名:多个输入时按顺序填入,其顺序与后续 onForward 中的输入数组需要保持一致
const std::vector<std::string> input_names{"input_1", "input_2", "input_3"};
// 输出名,可以为空,为空时 MNN 自动搜索模型中的输出,多输出情况下无法保证顺序,需要通过 getInfo 接口查看
// 输出名,多个输出按顺序填入,其顺序决定 onForward 的输出数组顺序
const std::vector<std::string> output_names{"output_1"};
Module::Config mdconfig; // default module config
Expand All @@ -56,6 +57,8 @@ std::unique_ptr<Module> module; // module
module.reset(Module::load(input_names, output_names, model_filename.c_str(), rtMgr, &mdconfig));
```

输入输出的名字可以为空,此时,MNN 会检索模型中的输入/输出填入,在多输入/输出情况下无法保证顺序,需要通过 getInfo 接口查看。

### Module::Config
创建`Module`时可传入`Module::Config`,具体结构如下:

Expand Down
56 changes: 56 additions & 0 deletions docs/pymnn/expr.md
Original file line number Diff line number Diff line change
Expand Up @@ -190,6 +190,62 @@ array([0., 1., 2., 3.], dtype=float32)
```python
>>> expr.set_global_executor_config(2, 2, 1)
```

---
### `sync()`
MNN VARP同步,调用后可以保证改VARP计算完毕

返回:`None`

返回类型:`None`

示例:

```python
>>> mnn_var = expr.placeholder([2,2])
>>> mnn_var.sync()
```

---
### `set_device_ptr(device_ptr, memory_type)`
设置MNN VARP GPU内存地址,同时指定给定内存地址对应的内存类型(CUDA/OPENCL/OPENGL等),仅在MNN VARP有GPU内存时可用:

参数:
- `device_ptr:uint64_t` 整形内存指针地址
- `memory_type:int` 例如: 2->CUDA 3->OpenCL等, 详见include/MNN/MNNForwardType.h文件中MNNForwardType结构体

返回:`None`

返回类型:`None`

示例:

```python
>>> torch_tensor = torch.empty([1, 1000], dtype=torch.float16).cuda()
>>> mnn_var = expr.placeholder([2,2])
>>> mnn_var.set_device_ptr(torch_tensor.data_ptr() ,2)
```

---
### `copy_to_device_ptr(device_ptr, memory_type)`
拷贝MNN VARP GPU内存到指定内存地址, 同时指定给定内存地址对应的内存类型(CUDA/OPENCL/OPENGL等):

参数:
- `device_ptr:uint64_t` 整形内存指针地址
- `memory_type:int` 例如: 2->CUDA 3->OpenCL等, 详见include/MNN/MNNForwardType.h文件中MNNForwardType结构体

返回:`None`

返回类型:`None`

示例:

```python
>>> torch_tensor = torch.empty([1, 1000], dtype=torch.float16).cuda()
>>> mnn_var = expr.placeholder([2,2])
>>> mnn_var.copy_to_device_ptr(torch_tensor.data_ptr() ,2)
```

---
### `sign(x)`
返回输入值的符号,正数返回1,负数返回-1
Expand Down
9 changes: 8 additions & 1 deletion express/Executor.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -58,6 +58,10 @@ void Executor::setGlobalExecutorConfig(MNNForwardType type, const BackendConfig&
mAttr->firstType = std::make_pair(type, numberThread);
info.mode = Backend::Info::DIRECT;
info.numThread = numberThread;
if (MNN_FORWARD_METAL == type) {
// Close metal's defer encoder
info.numThread |= MNN_GPU_RECORD_OP;
}
info.user = (BackendConfig*)&config;
std::shared_ptr<Runtime> bn(creator->onCreate(info));
mRuntimes[mAttr->firstType] = bn;
Expand Down Expand Up @@ -257,6 +261,9 @@ void Executor::RuntimeManager::setHint(Interpreter::HintMode mode, int value) {
case Interpreter::STRICT_CHECK_MODEL:
mInside->checkNetBuffer = value > 0;
break;
case Interpreter::MEM_ALLOCATOR_TYPE:
mInside->modes.memoryAllocatorType = value;
break;
default:
break;
}
Expand Down Expand Up @@ -538,7 +545,7 @@ void Executor::_makeCache(const std::vector<EXPRP>& expr, bool forceCPU) {
quant->scale = TensorUtils::getDescribe(srcTensor)->quantAttr.get()->scale;
quant->zero = TensorUtils::getDescribe(srcTensor)->quantAttr.get()->zero;
}

TensorUtils::getDescribe(tensor.get())->index = (int)scheduleInfo.allTensors.size();
scheduleInfo.allTensors.emplace_back(tensor);
}
Expand Down
42 changes: 42 additions & 0 deletions express/Expr.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -545,6 +545,48 @@ void Variable::setName(const std::string& name) {
mFrom->setName(name);
}
}

bool Variable::setDevicePtr(const void* devicePtr, int memoryType) {
if (nullptr != mFrom->get()) {
MNN_ERROR("Can't setDevicePtr to no-input op\n");
return false;
}
informDirty();
MNN_ASSERT(TensorUtils::getDescribe(mFrom->inside()->mOutputTensors[0])->quantAttr == nullptr || TensorUtils::getDescribe(mFrom->inside()->mOutputTensors[0])->type == DataType_DT_FLOAT);
mFrom->mInside->mContentDirty = false;
// Clear host address, Don't malloc hostPtr afterwards
Utils::releaseMemoryForHostTensor(mFrom->inside()->mOutputTensors[0]);
return mFrom->inside()->mOutputTensors[0]->setDevicePtr(devicePtr, memoryType);
}

bool Variable::copyToDevicePtr(void* devicePtr, int memoryType) {
if (nullptr != mFrom->get()) {
MNN_ERROR("Can't copyToDevicePtr to no-input op\n");
return false;
}

auto inside = mFrom->inside();
auto originTensor = inside->mOutputTensors[mFromIndex];

auto bn = TensorUtils::getDescribe(originTensor)->getBackend();
if(bn == nullptr) {
MNN_ERROR("Error: Varp copyToDevicePtr can't find backend\n");
return false;
}
if (bn->type() != memoryType) {
MNN_ERROR("Error: VARP backend type ( %d ), is not same as assigned memory type ( %d )\n", bn->type(), memoryType);
return false;
}

MNN::Tensor tempTensor(originTensor->dimensions(), originTensor->getDimensionType());
tempTensor.buffer().device = (uint64_t)devicePtr;

TensorUtils::getDescribe(originTensor)->getBackend()->onCopyBuffer(originTensor, &tempTensor);
// Sync the result
tempTensor.wait(Tensor::MAP_TENSOR_READ, true);
return true;
}

const std::string& Variable::name() const {
return mFrom->outputName(mFromIndex);
}
Expand Down
5 changes: 1 addition & 4 deletions express/Utils.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -116,10 +116,7 @@ bool Utils::allocMemoryForHostTensor(Tensor* dest) {
if (TensorUtils::getDescribe(dest)->memoryType != Tensor::InsideDescribe::MEMORY_HOST) {
return false;
}
auto size = dest->size();
if (0 >= size) {
return false;
}
auto size = dest->usize();
dest->buffer().host = (uint8_t*)MNNMemoryAllocAlign(size, MNN_MEMORY_ALIGN_DEFAULT);
return dest->buffer().host != nullptr;
}
Expand Down
5 changes: 4 additions & 1 deletion express/module/PipelineModule.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -335,7 +335,7 @@ static std::vector<int> _collectNeededOps(const MNN::Net* net, const std::set<in
// 0: use, 1: no use
std::vector<int> opMask(net->oplists()->size());
::memset(opMask.data(), 0, opMask.size() * sizeof(int));

// Set Initial Status
for (auto v : outputIndexes) {
tensorMask[v] = 1;
Expand Down Expand Up @@ -638,6 +638,9 @@ Module* PipelineModule::load(const std::vector<std::string>& inputs, const std::
modRuntime.rt.first.begin()->second->setExternalFile(rtMgr->getInside()->mExternalFile);
modRuntime.rt.second->setExternalFile(rtMgr->getInside()->mExternalFile);
}
// set allocator type
modRuntime.rt.first.begin()->second->setAllocatorType(rtMgr->getInside()->modes.memoryAllocatorType);
modRuntime.rt.second->setAllocatorType(rtMgr->getInside()->modes.memoryAllocatorType);
}
auto& rt = modRuntime.rt;
auto firstRt = rt.first[modRuntime.compute.type];
Expand Down
20 changes: 10 additions & 10 deletions include/MNN/ImageProcess.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,7 @@ enum Wrap { CLAMP_TO_EDGE = 0, ZERO = 1, REPEAT = 2 };
* 2: Sample line and do format convert
* 3: Turn RGBA to float tensor, and do sub and normalize
*/

class MNN_PUBLIC ImageProcess {
public:
struct Inside;
Expand All @@ -62,7 +63,6 @@ class MNN_PUBLIC ImageProcess {
/** edge wrapper */
Wrap wrap = CLAMP_TO_EDGE;
};

public:
/**
* @brief create image process with given config for given tensor.
Expand All @@ -86,10 +86,10 @@ class MNN_PUBLIC ImageProcess {
static ImageProcess* create(const ImageFormat sourceFormat = RGBA, const ImageFormat destFormat = RGBA,
const float* means = nullptr, const int meanCount = 0, const float* normals = nullptr,
const int normalCount = 0, const Tensor* dstTensor = nullptr);

~ImageProcess();
static void destroy(ImageProcess* imageProcess);

/**
* @brief get affine transform matrix.
* @return affine transform matrix.
Expand All @@ -98,7 +98,7 @@ class MNN_PUBLIC ImageProcess {
return mTransform;
}
void setMatrix(const Matrix& matrix);

/**
* @brief convert source data to given tensor.
* @param source source data.
Expand All @@ -109,7 +109,7 @@ class MNN_PUBLIC ImageProcess {
* @return result code.
*/
ErrorCode convert(const uint8_t* source, int iw, int ih, int stride, Tensor* dest);

/**
* @brief convert source data to given tensor.
* @param source source data.
Expand All @@ -126,7 +126,7 @@ class MNN_PUBLIC ImageProcess {
*/
ErrorCode convert(const uint8_t* source, int iw, int ih, int stride, void* dest, int ow, int oh, int outputBpp = 0,
int outputStride = 0, halide_type_t type = halide_type_of<float>());

/**
* @brief create tensor with given data.
* @param w image width.
Expand All @@ -140,7 +140,7 @@ class MNN_PUBLIC ImageProcess {
return createImageTensor(halide_type_of<T>(), w, h, bpp, p);
}
static Tensor* createImageTensor(halide_type_t type, int w, int h, int bpp, void* p = nullptr);

/**
* @brief set padding value when wrap=ZERO.
* @param value padding value.
Expand All @@ -149,14 +149,14 @@ class MNN_PUBLIC ImageProcess {
void setPadding(uint8_t value) {
mPaddingValue = value;
}

/**
* @brief set to draw mode.
* @param void
* @return void.
*/
void setDraw();

/**
* @brief draw color to regions of img.
* @param img the image to draw.
Expand All @@ -179,4 +179,4 @@ class MNN_PUBLIC ImageProcess {
} // namespace CV
} // namespace MNN

#endif /* ImageProcess_hpp */
#endif /* MNN_ImageProcess_hpp */
2 changes: 1 addition & 1 deletion include/MNN/MNNDefine.h
Original file line number Diff line number Diff line change
Expand Up @@ -69,6 +69,6 @@ MNN_ERROR("Check failed: %s ==> %s\n", #success, #log); \
#define STR(x) STR_IMP(x)
#define MNN_VERSION_MAJOR 2
#define MNN_VERSION_MINOR 8
#define MNN_VERSION_PATCH 1
#define MNN_VERSION_PATCH 2
#define MNN_VERSION STR(MNN_VERSION_MAJOR) "." STR(MNN_VERSION_MINOR) "." STR(MNN_VERSION_PATCH)
#endif /* MNNDefine_h */
4 changes: 4 additions & 0 deletions include/MNN/Tensor.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -304,6 +304,10 @@ class MNN_PUBLIC Tensor {
* @param finish wait for command flush or finish
*/
int wait(MapType mtype, bool finish);
/**
* @brief set GPU tensor device ptr, and inform memory type
*/
bool setDevicePtr(const void* devicePtr, int memoryType);
private:
halide_buffer_t mBuffer;
struct InsideDescribe* mDescribe;
Expand Down
5 changes: 4 additions & 1 deletion include/MNN/expr/Expr.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -108,11 +108,14 @@ class MNN_PUBLIC Variable {
Dimensionformat order = NHWC;
INTS dim;
halide_type_t type;
int size;
size_t size;
void syncSize();
};
const std::string& name() const;
void setName(const std::string& name);
bool setDevicePtr(const void* devicePtr, int memoryType);
bool copyToDevicePtr(void* devicePtr, int memoryType);

std::pair<EXPRP, int> expr() const {
return std::make_pair(mFrom, mFromIndex);
}
Expand Down
2 changes: 1 addition & 1 deletion project/android/gradle/wrapper/gradle-wrapper.properties
Original file line number Diff line number Diff line change
Expand Up @@ -3,4 +3,4 @@ distributionBase=GRADLE_USER_HOME
distributionPath=wrapper/dists
zipStoreBase=GRADLE_USER_HOME
zipStorePath=wrapper/dists
distributionUrl=https\://services.gradle.org/distributions/gradle-4.6-all.zip
distributionUrl=http\://mtl-gradle-mirror.oss-cn-hangzhou.aliyuncs.com/gradle-4.6-all.zip
Loading

0 comments on commit 970b63f

Please sign in to comment.