arm mali：针对buffer的特殊调度策略

知乎上有一个问题关于为何TNN的性能在mali上好，回答者和某用户进行了深入的沟通：

问：能否分享一下arm mali在 buffer/image 性能差异的原因，和异构调度策略的大概思路？
答：驱动设计层面的问题，clBuffer没有这个问题，华为有一套私有的API可以解决cIImage（性能差异）的问题。后来经过探索，发现opencl通过cpu周期性调用clFlush强制刷新command queue可绕过（arm mali在image/buffer上性能差异的）此问题。
参考：https://www.zhihu.com/question/400955777

下面，将针对TNN arm mali的调度策略进行分析，其回答中提到了clFlush，根据opencl对clFlush的解释，有如下说明：

clFlush会触发对已入命令队列中的命令的执行状态，且该命令队列是与当前设备绑定的；
clFlush只保证所有已经入队的命令发射，但不保证clFlush执行完后，命令队列就能完成计算；
其返回的状态码有如下类型：
- CL_SUCCESS表示执行成功；
- CL_INVALID_COMMAND_QUEUE表示无效的命令队列；
- CL_OUT_OF_HOST_MEMORY表示主机端分配OpenCL资源失败。

任何带有阻塞（blocking）的命令，都会隐式地对命令队列执行clFlush操作，下面4个例子：

带有blocking_read=CL_TRUE的clEnqueueReadImage；
带有blocking_write=CL_TRUE的clEnqueueReadBuffer、clEnqueueWriteImage和clEnqueueWriteBuffer；
带有blocking_map=CL_TRUE的clEnqueueMapImage、clEnqueueMapBuffer；
clWaitForEvents。

以上内容有些绕，总结一下：只有阻塞地Enqueue操作和WaitForEvent会隐式地去clFlush命令队列（即将命令推到发射状态），而非阻塞的Enqueue则不保证。

如若有两个命令队列，且他们有执行上的依赖关系，即命令队列B依赖命令队列A的执行，想要让命令队列A的Event对象作为B的条件，那么就需要隐式或显式地调用clFlush或阻塞命令，以确保入队的A任务的发射（start）状态。

arm官方文档在flush上的说明

参考：https://developer.arm.com/documentation/100614/0314/Retuning-existing-OpenCL-code/Differences-between-desktop-based-architectures-and-Mali-GPUs/About-Mali-GPU-architectures?lang=en

Avoid application processor and GPU interactions in the middle of processing

Enqueue all the kernels first, and call clFinish() at the end if possible. Call clFlush() after one or more clEnqueueNDRange() calls, and call clFinish() before checking the final result.

Avoid blocking calls in the submission thread

Avoid clFinish() or clWaitForEvent() or any other blocking calls in the submission thread. If possible, wait for an asynchronous callback if you want to check the result while computations are in progress. Try double buffering, if you are using blocking operations in your submission thread.

Batching kernels submission

From version r17p0 onwards, the OpenCL driver batches kernels that are flushed together for submission to the hardware. Batching kernels can significantly reduce the runtime overheads and cache maintenance costs. For example, this reduction is useful when the application is accessing multiple sub-buffers created from a buffer imported using clImportMemoryARM in separate kernels. The application should flush kernels in groups as large as possible. When the GPU is idle though, reaching optimal performance requires the application to flush an initial batch of kernels early so that the GPU execution overlaps the queuing of further kernels.

Execution optimizations

•   If you use callbacks to prompt the processor to continue processing data resulting from the execution of a kernel, ensure that the callbacks are set before you flush the queue. If you do not do this, the callbacks might occur at the end of a larger batch of work, later than they might have based on actual completion of work.

Wondering when I should use clFlush or clFinish.

参考：https://community.khronos.org/t/wondering-when-i-should-use-clflush-or-clfinish/3157

tnn/device/opencl/acc/opencl_layer_acc.cc

https://github.com/Tencent/TNN/blob/4b9ffbecc22f5ea4ba6bc4fdacff85475a59d08d/source/tnn/device/opencl/acc/opencl_layer_acc.cc#L160
Status OpenCLLayerAcc::Forward(const std::vector<Blob *> &inputs, const std::vector<Blob *> &outputs) {
    int unit_idx = 0;
    for (auto execute_unit : execute_units_) {
        ret = RunKernel(execute_unit.ocl_kernel, execute_unit.global_work_size, execute_unit.local_work_size,
                        ocl_context_->CommandQueue(), op_name_);
        unit_idx++;
    }

    if (NeedFlush()) {
        ocl_context_->CommandQueue()->flush();
    }

    return TNN_OK;
}

bool OpenCLLayerAcc::NeedFlush() {
    // flush by magic number
    if (0 == ocl_context_->AddAndGetFlushCount() % 10) {
        return true;
    }
    return false;
}

OpenCLContext

// https://github.com/Tencent/TNN/blob/a315d2acfb327014721b308359a6d534470289ba/source/tnn/device/opencl/opencl_context.cc
// opencl kernel flush strategy, some devices(special for huawei device) whave serious latency.
unsigned int OpenCLContext::AddAndGetFlushCount() {
    flush_count_++;
    return flush_count_;
}

// https://github.com/Tencent/TNN/blob/a315d2acfb327014721b308359a6d534470289ba/source/tnn/device/opencl/opencl_context.h#L88
class OpenCLContext : public Context {
public:
    OpenCLContext();
    ~OpenCLContext();

    // @brief get tnn command queue
    // @param command_queue device command queue for forward
    Status GetCommandQueue(void **command_queue) override;

    // @brief share tnn command queue to another context
    Status ShareCommandQueue(Context* context) override;

    /**
     * @brief get CommandQueue
     */
    cl::CommandQueue *CommandQueue();

    cl::CommandQueue *TuneCommandQueue();

    // load library
    virtual Status LoadLibrary(std::vector<std::string> path) override;
    /**
     * @brief befor instace forword
     * @param instance instace
     */
    virtual Status OnInstanceForwardBegin() override;
    /**
     * @brief after instace forword
     * @param instance instace
     */
    virtual Status OnInstanceForwardEnd() override;

     // @brief before instance Reshape
    virtual Status OnInstanceReshapeBegin() override;

    // @brief after instace Reshape
    virtual Status OnInstanceReshapeEnd() override;   

    // @brief wait for jobs in the current context to complete
    virtual Status Synchronize() override;

    // @brief add flush_count_ and return val
    unsigned int AddAndGetFlushCount();

    std::map<std::string, std::vector<uint32_t>>& GetLocalSizeTuneMap();

    Status StoreLocalSizeTuneMap();

public:
    /**
     * @brief initialize opencl env
     */
    Status Init();

private:
    std::shared_ptr<cl::CommandQueue> command_queue_ = nullptr;
    std::shared_ptr<cl::CommandQueue> tune_command_queue_ = nullptr;
    std::shared_ptr<cl::CommandQueue> GetCommandQueue();
    OpenCLRuntime *opencl_runtime_ = nullptr;
    unsigned int flush_count_ = 0;
    cl_command_queue_properties properties_ = 0;

    bool ReadStatusCheck(std::ifstream& is);

    std::map<std::string, std::vector<uint32_t>> local_size_tune_map_;
    uint32_t tune_map_size_;

    static std::mutex s_mutex_;

};

ysh329 commented 3 years ago

magic number for workgroup

https://github.com/Tencent/TNN/blob/aedc6c849e711a6386a8d2cd4ebb0bc94c7b9285/source/tnn/device/opencl/opencl_runtime.cc#L341

//magic number
static std::map<int, int> AdrenoSubGroup{
    {640, 128}, {630, 128}, {616, 128}, {612, 64}, {610, 64}, {540, 32}, {530, 32},
    {512, 32},  {510, 32},  {509, 32},  {506, 32}, {505, 32}, {405, 32}, {330, 16},
};

//opencl 2.0 can get SubGroupSize.
uint32_t OpenCLRuntime::GetSubGroupSize(const cl::Kernel &kernel, const cl::NDRange &range) {
    uint32_t sub_group_size = 0;

    if (ADRENO == gpu_info_.type) {
#if CL_HPP_TARGET_OPENCL_VERSION >= 200 && CL_TARGET_OPENCL_VERSION >= 210 && defined(CL_HPP_USE_CL_SUB_GROUPS_KHR)
        cl_int cl_ret;
        sub_group_size = kernel.getSubGroupInfo<CL_KERNEL_MAX_SUB_GROUP_SIZE_FOR_NDRANGE>(*device_, range, &cl_ret);
        if (cl_ret != CL_SUCCESS) {
            CHECK_CL_SUCCESS(cl_ret)
            sub_group_size = 0;
        }
#else
        if (AdrenoSubGroup.find(gpu_info_.model_num) != AdrenoSubGroup.end()) {
            sub_group_size = AdrenoSubGroup[gpu_info_.model_num];
        }
#endif
    }

    return sub_group_size;
}

ysh329 commented 3 years ago

cl&&gl交互

默认编译不开启，需要设置CMake

// https://github.com/Tencent/TNN/blob/aedc6c849e711a6386a8d2cd4ebb0bc94c7b9285/source/tnn/device/opencl/opencl_runtime.cc#L341
#ifdef SHARING_MEM_WITH_OPENGL
#include <EGL/egl.h>
#endif

//Init will get platforms info, get devices info, create opencl context.
Status OpenCLRuntime::Init() {
// ....

#if defined(SHARING_MEM_WITH_OPENGL) && (CL_HPP_TARGET_OPENCL_VERSION >= 120)
        // create context from glcontext
        LOGI("Create special opencl context to share with OpenGL\n");
        LOGI("eglGetCurrentContext(): 0x%x\n", eglGetCurrentContext());
        cl_context_properties context_prop[] = {CL_GL_CONTEXT_KHR, (cl_context_properties)eglGetCurrentContext(),
                                                CL_EGL_DISPLAY_KHR, (cl_context_properties)eglGetCurrentDisplay(), 0};
        context_ = std::shared_ptr<cl::Context>(new cl::Context(*device_, context_prop, nullptr, nullptr, &err));

        if (err != CL_SUCCESS) {
            LOGE(
                "Create special opencl context falied, Create common opencl "
                "context then.\n");
            context_ = std::shared_ptr<cl::Context>(new cl::Context(*device_, nullptr, nullptr, nullptr, &err));
        }
#else
        LOGI("Create common opencl context\n");
        context_ = std::shared_ptr<cl::Context>(new cl::Context(*device_, nullptr, nullptr, nullptr, &err));
#endif

ysh329 commented 3 years ago

tune

// https://github.com/Tencent/TNN/blob/4b9ffbecc22f5ea4ba6bc4fdacff85475a59d08d/source/tnn/device/opencl/acc/opencl_layer_acc.cc#L160

Status OpenCLLayerAcc::Forward(const std::vector<Blob *> &inputs, const std::vector<Blob *> &outputs) {
#if defined(LOCAL_SIZE_FINE_TUNE) && TNN_PROFILE
    auto execute_unit_org                                 = execute_units_[0];
    auto max_wgs                                          = execute_unit_org.workgroupsize_max;
    std::vector<std::vector<uint32_t>> local_size_list_3d = {
        {16, 4, 1}, {8, 8, 1},   {4, 16, 1}, {2, 32, 1}, {1, 64, 1}, {2, 64, 1}, {4, 64, 1},
        {8, 64, 1}, {16, 64, 1}, {8, 64, 2}, {4, 64, 4}, {2, 64, 8}, {2, 64, 4}, {},
    };
    std::vector<std::vector<uint32_t>> local_size_list_2d = {
        {2, max_wgs / 2},   {4, max_wgs / 4},   {8, max_wgs / 8},
        {16, max_wgs / 16}, {max_wgs / 2, 2},   {max_wgs / 4, 4},
        {max_wgs / 8, 8},   {max_wgs / 16, 16}, {},
    };
    std::vector<uint32_t> local_size_default;
    if (execute_unit_org.global_work_size.size() == 2) {
        local_size_default = LocalWS2DDefault(execute_unit_org);
    } else if (execute_unit_org.global_work_size.size() == 3) {
        local_size_default = LocalWS3DDefault(execute_unit_org);
    }

    OpenCLExecuteUnit exec_unit_default = execute_unit_org;
    exec_unit_default.local_work_size   = local_size_default;
    execute_units_.push_back(exec_unit_default);

    if (execute_unit_org.global_work_size.size() == 2) {
        for (auto local_size : local_size_list_2d) {
            OpenCLExecuteUnit exec_unit_temp = execute_unit_org;
            exec_unit_temp.local_work_size   = local_size;
            execute_units_.push_back(exec_unit_temp);
        }
    } else if (execute_unit_org.global_work_size.size() == 3) {
        for (auto local_size : local_size_list_3d) {
            OpenCLExecuteUnit exec_unit_temp = execute_unit_org;
            exec_unit_temp.local_work_size   = local_size;
            execute_units_.push_back(exec_unit_temp);
        }
    }

#endif

ysh329 / OpenCL-101

TNN对mali的调度、cl&gl交互、subworkgroup的magic number #57

arm mali：针对buffer的特殊调度策略

arm官方文档在flush上的说明

Avoid application processor and GPU interactions in the middle of processing

Avoid blocking calls in the submission thread

Batching kernels submission

Wondering when I should use clFlush or clFinish.

tnn/device/opencl/acc/opencl_layer_acc.cc

OpenCLContext

magic number for workgroup

cl&&gl交互

tune