还有， 我在vc中设了 /fp:fast 选项，但是好像效果不明显，您也可以试下。

这个选项对应numba的fastmath=True



【 在 ble 的大作中提到: 】
: 实际上你的C++都在做double加法，没法启用AVX；而Python代码都在做int加法，所以可以启用。
: #include <chrono>
: #include <format>
: ...................
--
FROM 123.112.71.*

这种简单代码，直接看生成的汇编就行

否则只是雾里看花
--
FROM 114.241.225.*

单线程跟多线程比没啥意义吧  
  
【 在 haha103 (haha103) 的大作中提到: 】
:  楼上的代码，稍微改了一下openmp线程数=16
:  
:  $ g++ -DNTHREADS=16 -std=c++20 -Wall -O3 -ffast-math -fopenmp -mavx2 -o out-avx2 test.cxx
:  $ ./out-avx2  
--
FROM 39.144.39.*

又换了台机器，objdump上传了，有时间的可以看看，哈哈。总的来说就是avx512比较上头，优势有点明显

给run0-6前面加了个__attribute__((noinline))方便大家看汇编


$ for arch in sse4.2 avx avx2 avx512f; do g++ -DNTHREADS=16 -std=c++20 -O3
-ffast-math -fopenmp -m$arch -o out-$arch test.cxx; done

$ for arch in sse4.2 avx avx2 avx512f; do echo "running $arch ..."; ./out-$arch; done
running sse4.2 ...
OpenMP init.
Run0 resut is 4999999950000000.00, done in 0.223278 s!
Run1 resut is 4999999950000000.00, done in 0.023822 s!
Run2 resut is 4999999950000000.00, done in 0.023822 s!
Run3 resut is 4999999950000000.00, done in 0.023821 s!
Run4 resut is 4999999950000000.00, done in 0.023828 s!
Run5 resut is 4999999950000000.00, done in 0.003083 s!
Run6 resut is 4999999950000000.00, done in 0.001811 s!
running avx ...
OpenMP init.
Run0 resut is 4999999950000000.00, done in 0.233588 s!
Run1 resut is 4999999950000000.00, done in 0.023821 s!
Run2 resut is 4999999950000000.00, done in 0.023824 s!
Run3 resut is 4999999950000000.00, done in 0.023822 s!
Run4 resut is 4999999950000000.00, done in 0.023822 s!
Run5 resut is 4999999950000000.00, done in 0.003914 s!
Run6 resut is 4999999950000000.00, done in 0.001868 s!
running avx2 ...
OpenMP init.
Run0 resut is 4999999950000000.00, done in 0.236559 s!
Run1 resut is 4999999950000000.00, done in 0.012901 s!
Run2 resut is 4999999950000000.00, done in 0.012878 s!
Run3 resut is 4999999950000000.00, done in 0.012875 s!
Run4 resut is 4999999950000000.00, done in 0.012875 s!
Run5 resut is 4999999950000000.00, done in 0.002789 s!
Run6 resut is 4999999950000000.00, done in 0.000998 s!
running avx512f ...
OpenMP init.
Run0 resut is 4999999950000000.00, done in 0.229596 s!
Run1 resut is 4999999950000000.00, done in 0.007151 s!
Run2 resut is 4999999950000000.00, done in 0.007041 s!
Run3 resut is 4999999950000000.00, done in 0.007238 s!
Run4 resut is 4999999950000000.00, done in 0.007241 s!
Run5 resut is 4999999950000000.00, done in 0.001357 s!
Run6 resut is 4999999950000000.00, done in 0.000531 s!

【 在 haha103 的大作中提到: 】
: 楼上的代码，稍微改了一下openmp线程数=16
: $ g++ -DNTHREADS=16 -std=c++20 -Wall -O3 -ffast-math -fopenmp -mavx2 -o out-avx2 test.cxx
: $ ./out-avx2
: ...................
--
修改:haha103 FROM 61.157.65.*
FROM 61.157.65.*
附件(328KB) out-avx.dump.gz
附件(329.8KB) out-avx2.dump.gz
附件(335.3KB) out-avx512f.dump.gz
附件(327.1KB) out-sse4.2.dump.gz

有可能是误差，这个结果不是很精确，毕竟那个机器还在干别的事情
【 在 Quanm 的大作中提到: 】
:     是什么东西的优化让 run5 的浮点数运算跟 run6 的整数运算速度差别不大的？
--
FROM 61.157.65.*