ranges::sort(data);  Eliminate mismatching begin/end iterators  Allows “range adaptors” to lazily transform/filter underlying sequences of elements12 Ranges  Based on following major components:  Range: A iterator pairs  Projection: Transform elements before handing over to algorithm  Views: transform/filter range: lazily evaluated, non-owning, non- mutating  Range factories: construct views to produce of chaining views: vector data { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 }; auto result { data | views::filter([](const auto& value) { return value % 2 == 0; })/* 2 4 6 8 10 */ | views::transform([](const

known.Memory model – Reference Counted Memory lifetime is tracked explicitly through reference counting. This typically means allocation locations are static. // 0x5200, ref 1 obj = create(); obj // this far more complicated since they are typically non- deterministic – possible even if Reference Counting is used. Consider accepting alloc/dealloc callbacks.You can make interop suck less by… Explicitly

as needed if the indicator is set. • Problem: Where is user code? In User Code • Integrated link counting. • Automatically repeat global cleanup if dependent retirement happens. • Implication: Asynchronous Global cleanup? • As in TS2? Depends on TS2 feedback. • Transitive cleanup after integration of link counting (as in Folly)? • Maybe deprecate global cleanup in Folly? • Other extensions: • Quasi-synchronous

for_each 的 lambda 没有修饰，而用于 x_dev 的那个 lambda 需要修饰 __device__ 。 for_each 用于整数的循环： counting_iterator • 可以用 thrust::make_counting_iterator(num) 构建一个计数迭代器，他作为区间表示的就是 整数的区间。 合并多个迭代器为一个： zip_iterator • 可以用

atomic>, atomic> • GH-601 by AdamBucior • (bit_cast, rotating/counting, power-of-2) • Several PRs by barcharcraz • make_shared() For Arrays • GH-309 by AdamBucior,

atomic ref_count … std::atomic ctrl Introduce a second layer of reference counting on the atomic instance itself! struct counted_cb { control_block* ctrl; long local_ref_count;

it’s safe to pass locals by reference to callees... … no dynamic allocation or reference counting needed.64 SENDER/RECEIVER IS ALSO STRUCTURED CONCURRENCY then then then schedule schedule

的 CTAD 特 性。第二个则是用了 decltype 自动推导 ，也可以 (auto r) ，这里写具体类型仅为 教学目的。 TBB 中其他并发容器 第 7 章：并行筛选 筛选（ filter ） 利用 vector 的 push_back 动态追加数据 筛选出所有大于 0 的 sin(i) 值 并行筛选 1 （张心欣犯过的错） 利用多线程安全的 concurrent_vector step 后依然写回了数组，数据缓存没法充分利用 。 另辟蹊径：流水线并行 加速比： 6.73 倍 反直觉的并行方式，但是加速效果却很理想，为什么？ 流水线模式下每个线程都只做自己的那个步骤（ filter ），从 而对指令缓存更友好。且一个核心处理完的数据很快会被另一 个核心用上，对三级缓存比较友好，也节省内存。 且 TBB 的流水线，其实比教科书上描述的传统流水线并行更加优化： 他在 t1 线程算完 能，完全无需操心内部 细节。 流水线并行： filter 参数 • serial_in_order 表示当前步骤只允许串行执行，且 执行的顺序必须一致。 • serial_out_of_order 表示只允许串行执行，但是顺 序可以打乱。 • parallel 表示可以并行执行当前步骤，且顺序可以 打乱。 • 每一个步骤（ filter ）的输入和返回类型都可以不一 样。要求：流水线上一步的返回类型，必须和下一

"as": "timestamp", "with": "%e-%b-%Y %H:%M", "tz": "Asia/Shanghai"} } } } {"swap": ["a", "b"] } {"filter": {"val1": 0 }} {"pivot_wider": {"names_from": "field1", "values_from": "field2"}} {"pivot_longer": P\d+)" } {"map": {"new1": { "expr": "field1.contains(\"abc\")", "as": "bool"}}} • Mutate • Filter • Pivot Powered by serde + arrow taosX - 数据汇聚 CONTENTS 自 我 介 绍 T D e n g i n e t a o s X

（ Property Graph KV store ） Pushdown filter Cypher (a)-[:LIKES]->(b) 计算下推：面向图应用特征设计 分布式存储 Find Vertex Walk Vertex Scan Cartesian Product Projection Filter Stage Stage 物理执行计划 01 02 03 将不同的执行阶段推送到对应的存储