modern CPU instruction cache. As a result, the hardware spends a considerable amount of processing time — nearly 30 percent, in many cases — getting an instruction stream from memory to the CPU.” Accelerate refactoring. 33Quick benchmark https://quick-bench.com/q/mOAHh7zZeagJlCJ63GtWVMPngsw 34 25% overhead {deferred_construction output; read_strings(in, out(output)); https://www.foonathan QWORD PTR [rdi]Quick benchmark https://quick-bench.com/q/pJ3z9L_Q1M16qob8-sg8lM3-T60 44 200% overhead {C++ Core Guidelines F.26: Use a unique_ptr to transfer ownership where a pointer is needed

1/66Bridging the Gap: Writing Portable Programs for CPU and GPU using CUDA Thomas Mejstrik Sebastian Woblistin 2/66Content 1 Motivation Audience etc.. Cuda crash course Quiz time 2 Patterns Oldschool afterwards7/66 Motivation Patterns The dark path Cuda proposal Thank you Why write programs for CPU and GPU Difference CPU/GPU Algorithms are designed differently Latency/Throughput Memory bandwidth Number of talk7/66 Motivation Patterns The dark path Cuda proposal Thank you Why write programs for CPU and GPU Difference CPU/GPU Why it makes sense? Library/Framework developers Embarrassingly parallel algorithms

Is std::mdspan a Zero-overhead Abstraction? Oleksandr Bacherikov Snap IncWhat is std::mdspan? It’s a view over a multi-dimensional array. It’s designed primarily to be used as a function parameter 24 Wrong! std::layout_stride supports only all strides specified at runtime. If we target zero overhead, we have to specify one of the strides as 1 at compile time.What does the Standard offer us instead

interpreter a JiT, or AoT? What if it modifies its bytecode? A CPU executes machine code… an interpreter executes bytecode… That’s the same thing, a CPU is an interpreter for machine code. A compiler can perform factor of 7 to 20. How do you develop hardware that doesn’t exist yet? When I started working on a CPU, it wasn’t obvious to me that you do so with simulators—multiple simulators—which simulate what the to the performance level of -O4.” “ Rosenblum et al. 1995 By far the fastest simulator of the CPU, MMU, and memory system of an SGI multiprocessor is an SGI multiprocessor. In other words, when

编译器如何自动优化：从汇编角度看 C++ 5.C++11 起的多线程编程：从 mutex 到无锁并行 6.并行编程常用框架： OpenMP 与 Intel TBB 7.被忽视的访存优化：内存带宽与 cpu 缓存机制 8.GPU 专题： wrap 调度，共享内存， barrier 9.并行算法实战： reduce ， scan ，矩阵乘法等 10.存储大规模三维数据的关键：稀疏数据结构 11.物理仿真实战：邻居搜索表实现 指望靠单核性能的增长带来程序性 能提升的时代一去不复返了，现在 要我们动动手为多核优化一下老的 程序，才能搭上摩尔定律的顺风车 。 神话与现实： 2 * 3GHz < 6GHz • 一个由双核组成的 3GHz 的 CPU 实际上提供了 6GHz 的处理能力，是吗？ • 显然不是。甚至在两个处理器上同时运行两个线程也不见得可以获得两倍的性能。相似的 ，大多数多线程的应用不会比双核处理器的两倍快。他们应该比单核处理器运行的快，但 是性能毕竟不是线性增长。 • 为什么无法做到呢？首先，为了保证缓存一致性以及其他握手协议需要运行时间开销。在 今天，双核或者四核机器在多线程应用方面，其性能不见得的是单核机器的两倍或者四倍。 这一问题一直伴随 CPU 发展至今。 并发和并行的区别 • 运用多线程的方式和动机，一般分为两种。 • 并发：单核处理器，操作系统通过时间片调 度算法，轮换着执行着不同的线程，看起来 就好像是同时运行一样，其实每一时刻只有

resolution) Negligible overhead (a few nanoseconds per zone) Cross-platform (Windows, Linux, macOS, iOS, Android, WASM*) Hybrid profiling capabilities (sampling and/or instrumentation) (CPU and GPU instrumentation) (analyze application while it is running) Precise measurements (nanosecond resolution) Negligible overhead (a few nanoseconds per zone) Cross-platform (Windows, Linux, macOS, iOS, Android, WASM*) Hybrid (analyze application while it is running) Precise measurements (nanosecond resolution) Negligible overhead (a few nanoseconds per zone) Cross-platform (Windows, Linux, macOS, iOS, Android, WASM*) Hybrid

resolution) Negligible overhead (a few nanoseconds per zone) Cross-platform (Windows, Linux, macOS, iOS, Android, WASM*) Hybrid profiling capabilities (sampling and/or instrumentation) (CPU and GPU instrumentation) (analyze application while it is running) Precise measurements (nanosecond resolution) Negligible overhead (a few nanoseconds per zone) Cross-platform (Windows, Linux, macOS, iOS, Android, WASM*) Hybrid (analyze application while it is running) Precise measurements (nanosecond resolution) Negligible overhead (a few nanoseconds per zone) Cross-platform (Windows, Linux, macOS, iOS, Android, WASM*) Hybrid

creation is CPU and memory intensive19/55 Threads - Evaluation Thread creation is CPU and memory intensive Thread switching is expensive19/55 Threads - Evaluation Thread creation is CPU and memory threads may still consume CPU cycles!19/55 Threads - Evaluation Thread creation is CPU and memory intensive Thread switching is expensive Blocked threads may still consume CPU cycles! Shared data must correctly.19/55 Threads - Evaluation Thread creation is CPU and memory intensive Thread switching is expensive Blocked threads may still consume CPU cycles! Shared data must be synchronized correctly.

lock unlock unlock unlockthreads (visualized) CPU A CPU B CPU C lock lock unlock unlock unlockactors (visualized) CPU A CPU B request CPU C request response responseactor performance for different thread than the one we started on, we’d be surprised! ● ... on a GPU when we started on a CPU, we’d be surprised!function abstraction std::string SpellCheck(std::string text) { auto body = different thread than the one we started on, we’d be surprised! ● ... on a GPU when we started on a CPU, we’d be surprised! breaks the principle of least astonishment!where should the continuation run?

available for customization. 4344 44 JOB SYSTEM • Scheduler + User defined workload • Optimize for CPU throughput Okay, finally, we are ready to talk about the job system! What is a job system? Job system bigger overall CPU throughput, which means ideally, I will want all threads be busy and have things to do all the times, and we want to avoid bubbles, which is spared time frames on the CPU timeline. We store a future pointer. Notice here, future is not the std::future, std::future has too much overhead for our usage here, this is a utility type provided in the system. 8182 future template

0 码力 | 120 页 | 2.20 MB | 6 月前

3