pointers protect access to objects that may be removed concurrently. SAFE RECLAMATION Concurrency TS2 Essential Hazard Pointer Interface Base class for protectable objects template class time) Asynchronous Reclamation • Asynchronous reclamation is invoked when the number of retired objects reaches some threshold: • In the Folly library: • The threshold is the max of 1000 and twice the Extract retired objects from lists in the (global) domain structure. • Read hazard pointer values • Match addresses of retired objects with values read from hazard pointers. • Push matched objects back into

N*sizeof(float)); // … cudaFree(x); cudaFree(y); } An Even Easier Introduction to CUDA 5 |__global__ void add_gpu(int n, float* x, float* y) { for (int i = 0; i < n; i++) y[i] = x[i] y[i]; } TEST_CASE("cppcon-1", "[CUDA]") { // … } An Even Easier Introduction to CUDA 6 |__global__ void add_gpu(int n, float* x, float* y) { for (int i = 0; i < n; i++) y[i] = x[i] + // … add_gpu<<<1, 1>>>(N, x, y); // … } An Even Easier Introduction to CUDA 7 |__global__ void add_gpu(int n, float* x, float* y) { for (int i = 0; i < n; i++) y[i] = x[i]

protect access to smart pointer  Use global non-member atomic operations, e.g. std::atomic_load(), atomic_store(), …  Error-prone, easy to accidently not use global non-member atomic operations  C++20: C++20: atomic>  Might use mutex internally  Global non-member atomic operations are deprecated36 Atomic Smart Pointers template class concurrent_stack { struct Node { T Spaceship Operator <=>  Official name: three-way comparison operator  Three-way: comparing 2 objects and then comparing result with 0  (a <=> b) < 0 // true if a < b  (a <=> b) > 0 // true if

wrapper headers (e.g. std::fopen())  import std.compat;  Imports everything std imports  + global namespace versions of the C wrapper headers (e.g. ::fopen()) Standard Library Modules42  Modern  We already have heterogeneous lookup for associative containers  Avoids creating temporary objects of type key during lookups  E.g.: lookup with C-style string for container with std::string as

C++ 头文件等。 • host 代码和 device 代码写在同一个文件内，这 是 OpenCL 做不到的。 编写一段在 GPU 上运行的代码 • 定义函数 kernel ，前面加上 __global__ 修 饰符，即可让他在 GPU 上执行。 • 不过调用 kernel 时，不能直接 kernel() ，而 是要用 kernel<<<1, 1>>>() 这样的三重尖括 号语法。为什么？这里面的两个 号语法。为什么？这里面的两个 1 有什么用 ？稍后会说明。 • 运行以后，就会在 GPU 上执行 printf 了。 • 这里的 kernel 函数在 GPU 上执行，称为核 函数，用 __global__ 修饰的就是核函数。 没有反应？同步一下！ • 然而如果直接编译运行刚刚那段代码，是不会打印出 Hello, world! 的。 • 这是因为 GPU 和 CPU 之间的通信，为了高效，是异 上的设备函数 • __global__ 用于定义核函数，他在 GPU 上执行，从 CPU 端通过三重尖括号语法调 用，可以有参数，不可以有返回值。 • 而 __device__ 则用于定义设备函数，他在 GPU 上执行，但是从 GPU 上调用的，而 且不需要三重尖括号，和普通函数用起来一 样，可以有参数，有返回值。 • 即： host 可以调用 global ； global 可以调 用

spawn_blocking 调度模式 spawn 调度模式 Thread Worker task Local queue Thread Thread task Global queue task New task Global queue New task take & run take & run Worker take & run Steal & run 两种接口拥有两套割裂的调度模式和线程池 高优先级任务由高权重线程调度， 以此获得更多执行时间 • 全局队列区分高低优先级 Task priority and quality of service 高权重线程 低权重线程 task …. task …. Global queue task Local queue task Local queue Core 高权重线程 Worker Worker 任务优先级调度 根据工作线程的优先级进行绑核（大小核）

start() places a callback that completes the receiver in a global. 3. Register a keyboard callback that reads the completion info out of the global and completes it if it’s not null.86 KEYCLICK SENDER pending_completion { virtual void complete(char) = 0; virtual ~pending_completion() {} }; // Global registration of next completion: std::atomic pending_completion_{nullptr};

a tzdb • There also exists a type called tzdb_list which contains a list of tzdb objects. • Since tzdb objects can be tied to newer versions of the time zone data, this list can contain various

must have moved my local count to the global count } } The store operation helps an in- flight load by moving its local_ref_count onto the global ref_count30 Daniel Anderson -- danielanderson

being on the same thread. Too much work on main thread • Android nested layouts • Functions and objects defined in loops • Statements like debugger, eval, with. • How to access Native Engine information