【发布时间】:2016-02-14 02:31:02
【问题描述】:
披露:我在programmers.stack 上尝试过类似的问题,但那个地方离活动堆栈不远。
简介
我倾向于处理大量大图像。它们还以不止一个序列出现,并且必须反复处理和播放。有时我使用 GPU,有时使用 CPU,有时两者兼而有之。大多数访问模式本质上是线性的(来回),这让我开始思考关于数组的更基本的事情,以及一种方法应该如何编写针对给定硬件上可能的最大内存带宽优化的代码(允许计算不会阻塞读/写) .
测试规范
- 我在 2011 MacbookAir4,2 (I5-2557M) 上完成了这项工作,配备 4GB RAM 和 SSD。测试期间除了 iterm2 之外没有其他任何东西在运行。
- gcc 5.2.0 (homebrew) 带有标志:
-pedantic -std=c99 -Wall -Werror -Wextra -Wno-unused -O0带有额外的包含和库标志以及框架标志,以便使用我倾向于使用的 glfw 计时器。我本来可以不做的,没关系。当然,全是 64 位。 - 我已经尝试使用可选的
-fprefetch-loop-arrays标志进行测试,但它似乎根本没有影响结果
测试
- 在堆上分配两个
n bytes数组 - 其中n是8, 16, 32, 64, 128, 256, 512 and 1024 MB - 将
array初始化为0xff,一次一个字节 - 测试 1 - 线性复制
线性复制:
for(uint64_t i = 0; i < ARRAY_NUM; ++i) {
array_copy[i] = array[i];
}
-
测试 2 - 大步复制。这就是令人困惑的地方。我试过在这里玩预取游戏。我已经尝试了每个循环我应该做多少的各种组合,似乎每个循环约 40 次会产生最佳性能。 为什么?我不知道。我知道 c99 中的
malloc和uint64_t会给我内存对齐的块。我还看到我的 L1 到 L3 缓存的大小高于这些320 bytes,那么我在打什么?线索可能稍后出现在图表中。我真的很想了解这一点。
跨步复制:
for(uint64_t i = 0; i < ARRAY_NUM; i=i+40) {
array_copy[i] = array[i];
array_copy[i+1] = array[i+1];
array_copy[i+2] = array[i+2];
array_copy[i+3] = array[i+3];
array_copy[i+4] = array[i+4];
array_copy[i+5] = array[i+5];
array_copy[i+6] = array[i+6];
array_copy[i+7] = array[i+7];
array_copy[i+8] = array[i+8];
array_copy[i+9] = array[i+9];
array_copy[i+10] = array[i+10];
array_copy[i+11] = array[i+11];
array_copy[i+12] = array[i+12];
array_copy[i+13] = array[i+13];
array_copy[i+14] = array[i+14];
array_copy[i+15] = array[i+15];
array_copy[i+16] = array[i+16];
array_copy[i+17] = array[i+17];
array_copy[i+18] = array[i+18];
array_copy[i+19] = array[i+19];
array_copy[i+20] = array[i+20];
array_copy[i+21] = array[i+21];
array_copy[i+22] = array[i+22];
array_copy[i+23] = array[i+23];
array_copy[i+24] = array[i+24];
array_copy[i+25] = array[i+25];
array_copy[i+26] = array[i+26];
array_copy[i+27] = array[i+27];
array_copy[i+28] = array[i+28];
array_copy[i+29] = array[i+29];
array_copy[i+30] = array[i+30];
array_copy[i+31] = array[i+31];
array_copy[i+32] = array[i+32];
array_copy[i+33] = array[i+33];
array_copy[i+34] = array[i+34];
array_copy[i+35] = array[i+35];
array_copy[i+36] = array[i+36];
array_copy[i+37] = array[i+37];
array_copy[i+38] = array[i+38];
array_copy[i+39] = array[i+39];
}
- 测试 3 - 大步阅读。与大步复制相同。
大步阅读:
const int imax = 1000;
for(int j = 0; j < imax; ++j) {
uint64_t tmp = 0;
performance = 0;
time_start = glfwGetTime();
for(uint64_t i = 0; i < ARRAY_NUM; i=i+40) {
tmp = array[i];
tmp = array[i+1];
tmp = array[i+2];
tmp = array[i+3];
tmp = array[i+4];
tmp = array[i+5];
tmp = array[i+6];
tmp = array[i+7];
tmp = array[i+8];
tmp = array[i+9];
tmp = array[i+10];
tmp = array[i+11];
tmp = array[i+12];
tmp = array[i+13];
tmp = array[i+14];
tmp = array[i+15];
tmp = array[i+16];
tmp = array[i+17];
tmp = array[i+18];
tmp = array[i+19];
tmp = array[i+20];
tmp = array[i+21];
tmp = array[i+22];
tmp = array[i+23];
tmp = array[i+24];
tmp = array[i+25];
tmp = array[i+26];
tmp = array[i+27];
tmp = array[i+28];
tmp = array[i+29];
tmp = array[i+30];
tmp = array[i+31];
tmp = array[i+32];
tmp = array[i+33];
tmp = array[i+34];
tmp = array[i+35];
tmp = array[i+36];
tmp = array[i+37];
tmp = array[i+38];
tmp = array[i+39];
}
-
测试 4 - 线性读数。每个字节一个字节。我很惊讶
-fprefetch-loop-arrays在这里没有产生任何结果。我以为是针对这些情况的。
线性读数:
for(uint64_t i = 0; i < ARRAY_NUM; ++i) {
tmp = array[i];
}
-
测试 5 -
memcpy作为对比。
memcpy:
memcpy(array_copy, array, ARRAY_NUM*sizeof(uint64_t));
结果
- 样本输出:
样本输出:
Init done in 0.767 s - size of array: 1024 MBs (x2)
Performance: 1304.325 MB/s
Copying (linear) done in 0.898 s
Performance: 1113.529 MB/s
Copying (stride 40) done in 0.257 s
Performance: 3890.608 MB/s
[1000/1000] Performance stride 40: 7474.322 MB/s
Average: 7523.427 MB/s
Performance MIN: 3231 MB/s | Performance MAX: 7818 MB/s
[1000/1000] Performance dumb: 2504.713 MB/s
Average: 2481.502 MB/s
Performance MIN: 1572 MB/s | Performance MAX: 2644 MB/s
Copying (memcpy) done in 1.726 s
Performance: 579.485 MB/s
--
Init done in 0.415 s - size of array: 512 MBs (x2)
Performance: 1233.136 MB/s
Copying (linear) done in 0.442 s
Performance: 1157.147 MB/s
Copying (stride 40) done in 0.116 s
Performance: 4399.606 MB/s
[1000/1000] Performance stride 40: 6527.004 MB/s
Average: 7166.458 MB/s
Performance MIN: 4359 MB/s | Performance MAX: 7787 MB/s
[1000/1000] Performance dumb: 2383.292 MB/s
Average: 2409.005 MB/s
Performance MIN: 1673 MB/s | Performance MAX: 2641 MB/s
Copying (memcpy) done in 0.102 s
Performance: 5026.476 MB/s
--
Init done in 0.228 s - size of array: 256 MBs (x2)
Performance: 1124.618 MB/s
Copying (linear) done in 0.242 s
Performance: 1057.916 MB/s
Copying (stride 40) done in 0.070 s
Performance: 3650.996 MB/s
[1000/1000] Performance stride 40: 7129.206 MB/s
Average: 7370.537 MB/s
Performance MIN: 4805 MB/s | Performance MAX: 7848 MB/s
[1000/1000] Performance dumb: 2456.129 MB/s
Average: 2435.556 MB/s
Performance MIN: 1496 MB/s | Performance MAX: 2637 MB/s
Copying (memcpy) done in 0.050 s
Performance: 5095.845 MB/s
--
Init done in 0.100 s - size of array: 128 MBs (x2)
Performance: 1277.200 MB/s
Copying (linear) done in 0.112 s
Performance: 1147.030 MB/s
Copying (stride 40) done in 0.029 s
Performance: 4424.513 MB/s
[1000/1000] Performance stride 40: 6497.635 MB/s
Average: 6714.540 MB/s
Performance MIN: 4206 MB/s | Performance MAX: 7843 MB/s
[1000/1000] Performance dumb: 2275.336 MB/s
Average: 2335.544 MB/s
Performance MIN: 1572 MB/s | Performance MAX: 2626 MB/s
Copying (memcpy) done in 0.025 s
Performance: 5086.502 MB/s
--
Init done in 0.051 s - size of array: 64 MBs (x2)
Performance: 1255.969 MB/s
Copying (linear) done in 0.058 s
Performance: 1104.282 MB/s
Copying (stride 40) done in 0.015 s
Performance: 4305.765 MB/s
[1000/1000] Performance stride 40: 7750.063 MB/s
Average: 7412.167 MB/s
Performance MIN: 3892 MB/s | Performance MAX: 7826 MB/s
[1000/1000] Performance dumb: 2610.136 MB/s
Average: 2577.313 MB/s
Performance MIN: 2126 MB/s | Performance MAX: 2652 MB/s
Copying (memcpy) done in 0.013 s
Performance: 4871.823 MB/s
--
Init done in 0.024 s - size of array: 32 MBs (x2)
Performance: 1306.738 MB/s
Copying (linear) done in 0.028 s
Performance: 1148.582 MB/s
Copying (stride 40) done in 0.008 s
Performance: 4265.907 MB/s
[1000/1000] Performance stride 40: 6181.040 MB/s
Average: 7124.592 MB/s
Performance MIN: 3480 MB/s | Performance MAX: 7777 MB/s
[1000/1000] Performance dumb: 2508.669 MB/s
Average: 2556.529 MB/s
Performance MIN: 1966 MB/s | Performance MAX: 2646 MB/s
Copying (memcpy) done in 0.007 s
Performance: 4617.860 MB/s
--
Init done in 0.013 s - size of array: 16 MBs (x2)
Performance: 1243.011 MB/s
Copying (linear) done in 0.014 s
Performance: 1139.362 MB/s
Copying (stride 40) done in 0.004 s
Performance: 4181.548 MB/s
[1000/1000] Performance stride 40: 6317.129 MB/s
Average: 7358.539 MB/s
Performance MIN: 5250 MB/s | Performance MAX: 7816 MB/s
[1000/1000] Performance dumb: 2529.707 MB/s
Average: 2525.783 MB/s
Performance MIN: 1823 MB/s | Performance MAX: 2634 MB/s
Copying (memcpy) done in 0.003 s
Performance: 5167.561 MB/s
--
Init done in 0.007 s - size of array: 8 MBs (x2)
Performance: 1186.019 MB/s
Copying (linear) done in 0.007 s
Performance: 1147.018 MB/s
Copying (stride 40) done in 0.002 s
Performance: 4157.658 MB/s
[1000/1000] Performance stride 40: 6958.839 MB/s
Average: 7097.742 MB/s
Performance MIN: 4278 MB/s | Performance MAX: 7499 MB/s
[1000/1000] Performance dumb: 2585.366 MB/s
Average: 2537.896 MB/s
Performance MIN: 2284 MB/s | Performance MAX: 2610 MB/s
Copying (memcpy) done in 0.002 s
Performance: 5059.164 MB/s
- 线性阅读比跨步阅读慢 3 倍。步幅读数最大值约为。 7500-7800 MB/s 范围。不过,有两件事让我感到困惑。在 DDR3 1333 Mhz 下,最大内存吞吐量应该是
10,664 MB/s那么为什么我没有达到呢?为什么阅读速度不一致,我将如何优化(缓存未命中?)?从图表中可以更明显地看出,尤其是性能经常下降的线性读数。
图表
这里是所有感兴趣的人的完整来源:
/*
gcc -pedantic -std=c99 -Wall -Werror -Wextra -Wno-unused -O0 -I "...path to glfw3 includes ..." -L "...path to glfw3 lib ..." arr_test_copy_gnuplot.c -o arr_test_copy_gnuplot -lglfw3 -framework OpenGL -framework Cocoa -framework IOKit -framework CoreVideo
optional: -fprefetch-loop-arrays
*/
#include <stdio.h>
#include <stdlib.h>
#include <string.h> /* memcpy */
#include <inttypes.h>
#include <GLFW/glfw3.h>
#define ARRAY_NUM 1000000 * 128 /* GIG */
int main(int argc, char *argv[]) {
if(!glfwInit()) {
exit(EXIT_FAILURE);
}
int cx = 0;
char filename_stride[50];
char filename_dumb[50];
cx = snprintf(filename_stride, 50, "%lu_stride.dat",
((ARRAY_NUM*sizeof(uint64_t))/1000000));
if(cx < 0 || cx >50) { exit(EXIT_FAILURE); }
FILE *file_stride = fopen(filename_stride, "w");
cx = snprintf(filename_dumb, 50, "%lu_dumb.dat",
((ARRAY_NUM*sizeof(uint64_t))/1000000));
if(cx < 0 || cx >50) { exit(EXIT_FAILURE); }
FILE *file_dumb = fopen(filename_dumb, "w");
if(file_stride == NULL || file_dumb == NULL) {
perror("Error opening file.");
exit(EXIT_FAILURE);
}
uint64_t *array = malloc(sizeof(uint64_t) * ARRAY_NUM);
uint64_t *array_copy = malloc(sizeof(uint64_t) * ARRAY_NUM);
double performance = 0.0;
double time_start = 0.0;
double time_end = 0.0;
double performance_min = 0.0;
double performance_max = 0.0;
/* Init array */
time_start = glfwGetTime();
for(uint64_t i = 0; i < ARRAY_NUM; ++i) {
array[i] = 0xff;
}
time_end = glfwGetTime();
performance = ((ARRAY_NUM * sizeof(uint64_t))/1000000) / (time_end - time_start);
printf("Init done in %.3f s - size of array: %lu MBs (x2)\n", (time_end - time_start), (ARRAY_NUM*sizeof(uint64_t)/1000000));
printf("Performance: %.3f MB/s\n\n", performance);
/* Linear copy */
performance = 0;
time_start = glfwGetTime();
for(uint64_t i = 0; i < ARRAY_NUM; ++i) {
array_copy[i] = array[i];
}
time_end = glfwGetTime();
performance = ((ARRAY_NUM * sizeof(uint64_t))/1000000) / (time_end - time_start);
printf("Copying (linear) done in %.3f s\n", (time_end - time_start));
printf("Performance: %.3f MB/s\n\n", performance);
/* Copying with wide stride */
performance = 0;
time_start = glfwGetTime();
for(uint64_t i = 0; i < ARRAY_NUM; i=i+40) {
array_copy[i] = array[i];
array_copy[i+1] = array[i+1];
array_copy[i+2] = array[i+2];
array_copy[i+3] = array[i+3];
array_copy[i+4] = array[i+4];
array_copy[i+5] = array[i+5];
array_copy[i+6] = array[i+6];
array_copy[i+7] = array[i+7];
array_copy[i+8] = array[i+8];
array_copy[i+9] = array[i+9];
array_copy[i+10] = array[i+10];
array_copy[i+11] = array[i+11];
array_copy[i+12] = array[i+12];
array_copy[i+13] = array[i+13];
array_copy[i+14] = array[i+14];
array_copy[i+15] = array[i+15];
array_copy[i+16] = array[i+16];
array_copy[i+17] = array[i+17];
array_copy[i+18] = array[i+18];
array_copy[i+19] = array[i+19];
array_copy[i+20] = array[i+20];
array_copy[i+21] = array[i+21];
array_copy[i+22] = array[i+22];
array_copy[i+23] = array[i+23];
array_copy[i+24] = array[i+24];
array_copy[i+25] = array[i+25];
array_copy[i+26] = array[i+26];
array_copy[i+27] = array[i+27];
array_copy[i+28] = array[i+28];
array_copy[i+29] = array[i+29];
array_copy[i+30] = array[i+30];
array_copy[i+31] = array[i+31];
array_copy[i+32] = array[i+32];
array_copy[i+33] = array[i+33];
array_copy[i+34] = array[i+34];
array_copy[i+35] = array[i+35];
array_copy[i+36] = array[i+36];
array_copy[i+37] = array[i+37];
array_copy[i+38] = array[i+38];
array_copy[i+39] = array[i+39];
}
time_end = glfwGetTime();
performance = ((ARRAY_NUM * sizeof(uint64_t))/1000000) / (time_end - time_start);
printf("Copying (stride 40) done in %.3f s\n", (time_end - time_start));
printf("Performance: %.3f MB/s\n\n", performance);
/* Reading with wide stride */
const int imax = 1000;
double performance_average = 0.0;
for(int j = 0; j < imax; ++j) {
uint64_t tmp = 0;
performance = 0;
time_start = glfwGetTime();
for(uint64_t i = 0; i < ARRAY_NUM; i=i+40) {
tmp = array[i];
tmp = array[i+1];
tmp = array[i+2];
tmp = array[i+3];
tmp = array[i+4];
tmp = array[i+5];
tmp = array[i+6];
tmp = array[i+7];
tmp = array[i+8];
tmp = array[i+9];
tmp = array[i+10];
tmp = array[i+11];
tmp = array[i+12];
tmp = array[i+13];
tmp = array[i+14];
tmp = array[i+15];
tmp = array[i+16];
tmp = array[i+17];
tmp = array[i+18];
tmp = array[i+19];
tmp = array[i+20];
tmp = array[i+21];
tmp = array[i+22];
tmp = array[i+23];
tmp = array[i+24];
tmp = array[i+25];
tmp = array[i+26];
tmp = array[i+27];
tmp = array[i+28];
tmp = array[i+29];
tmp = array[i+30];
tmp = array[i+31];
tmp = array[i+32];
tmp = array[i+33];
tmp = array[i+34];
tmp = array[i+35];
tmp = array[i+36];
tmp = array[i+37];
tmp = array[i+38];
tmp = array[i+39];
}
time_end = glfwGetTime();
performance = ((ARRAY_NUM * sizeof(uint64_t))/1000000) / (time_end - time_start);
performance_average += performance;
if(performance > performance_max) { performance_max = performance; }
if(j == 0) { performance_min = performance; }
if(performance < performance_min) { performance_min = performance; }
printf("[%d/%d] Performance stride 40: %.3f MB/s\r", j+1, imax, performance);
fprintf(file_stride, "%d\t%f\n", j, performance);
fflush(file_stride);
fflush(stdout);
}
performance_average = performance_average / imax;
printf("\nAverage: %.3f MB/s\n", performance_average);
printf("Performance MIN: %3.f MB/s | Performance MAX: %3.f MB/s\n\n",
performance_min, performance_max);
/* Linear reading */
performance_average = 0.0;
performance_min = 0.0;
performance_max = 0.0;
for(int j = 0; j < imax; ++j) {
uint64_t tmp = 0;
performance = 0;
time_start = glfwGetTime();
for(uint64_t i = 0; i < ARRAY_NUM; ++i) {
tmp = array[i];
}
time_end = glfwGetTime();
performance = ((ARRAY_NUM * sizeof(uint64_t))/1000000) / (time_end - time_start);
performance_average += performance;
if(performance > performance_max) { performance_max = performance; }
if(j == 0) { performance_min = performance; }
if(performance < performance_min) { performance_min = performance; }
printf("[%d/%d] Performance dumb: %.3f MB/s\r", j+1, imax, performance);
fprintf(file_dumb, "%d\t%f\n", j, performance);
fflush(file_dumb);
fflush(stdout);
}
performance_average = performance_average / imax;
printf("\nAverage: %.3f MB/s\n", performance_average);
printf("Performance MIN: %3.f MB/s | Performance MAX: %3.f MB/s\n\n",
performance_min, performance_max);
/* Memcpy */
performance = 0;
time_start = glfwGetTime();
memcpy(array_copy, array, ARRAY_NUM*sizeof(uint64_t));
time_end = glfwGetTime();
performance = ((ARRAY_NUM * sizeof(uint64_t))/1000000) / (time_end - time_start);
printf("Copying (memcpy) done in %.3f s\n", (time_end - time_start));
printf("Performance: %.3f MB/s\n", performance);
/* Cleanup and exit */
free(array);
free(array_copy);
glfwTerminate();
fclose(file_dumb);
fclose(file_stride);
exit(EXIT_SUCCESS);
}
总结
- 在处理线性访问是最常见模式的数组时,我应该如何编写代码以获得最大和(接近)恒定速度?
- 我可以从这个示例中了解有关缓存和预取的哪些信息?
- 这些图表是否告诉我一些我应该知道但我没有注意到的事情?
- 我还能如何展开循环?我试过
-funroll-loops没有结果,所以我求助于手动编写循环展开。
感谢您的长期阅读。
编辑:
似乎-O0 与-O 标志不存在时的表现不同!是什么赋予了?如图所示,不存在标志会产生更好的性能。
EDIT2:
我终于用 AVX 达到了顶峰。
=== READING WITH AVX ===
[1000/1000] Performance AVX: 9868.912 MB/s
Average: 10029.085 MB/s
Performance MIN: 6554 MB/s | Performance MAX: 11464 MB/s
平均值非常接近 10664。我不得不将编译器更改为 clang,因为 gcc 让我很难使用 avx (-mavx)。这也是为什么图表有更明显的下降。我仍然想知道如何/什么是/有稳定的表现。我认为这是由于缓存/缓存行。它还解释了性能高于 DDR3 速度的原因(MAX 为 11464 MB/s)。
请原谅我的 gnuplot-fu 及其键。蓝色是 SSE2 (_mm_load_si128),橙色是 AVX (_mm256_load_si256)。紫色像以前一样大步走,绿色是哑巴一次读一个。
所以,最后两个问题是:
- 是什么导致了低谷以及如何保持更稳定的性能
- 没有内在函数是否有可能达到天花板?
最新版本的要点:https://gist.github.com/Keyframe/1ed9062ec52fc4a0d14b 和该版本的图表:http://imgur.com/a/cPeor
【问题讨论】:
-
为什么编译时不优化?
-
在这种情况下,打开优化(-O3 或任何其他)会产生奇怪的结果,跳过代码等。完全关闭它们我希望能更清楚地了解发生了什么我专注于 asm 输出。
-
@Keyframe:没有优化,你就没有衡量任何值得一提的东西。使用
-O3和-march=native,gcc将在适当的时候使用 SIMD 指令(并非一直如此,因此 GCC 编译器内在函数有时对强制它很有用)。关闭优化后,它不会那样做,也不会有任何其他明显的改进。volatile目标也不是真正的答案,因为这可能会在某些架构上导致不必要的内存屏障,从而完全破坏执行加速的任何尝试。 -
我不想这么说,但没有优化的编译不是解决方案,“-O3 很奇怪”。如果优化正在绕过你的东西,你需要改变你的代码,使它不能。这意味着使用计算结果,以便编译器无法删除它。或者切换到不会进行更改访问模式的疯狂循环转换的较低级别的优化。当你在没有优化的情况下编译时,它可能会很慢,以至于它完全隐藏了内存访问中的任何差异。
-
GCC 很聪明,它会优化掉测试 3 和 4
标签: c performance memory c99