我正在uint64_t使用.RAW文件中的RGB数据对10+百万个数据进行排序,而我的C程序时间中有79%花费在上qsort。我正在为这种特定的数据类型寻找更快的排序方式。
uint64_t
qsort
作为RAW图形数据,数字是非常随机的,大约80%是唯一的。不能期望部分排序或运行排序后的数据。uint16_t里面的4 s uint64_t是R,G,B和零(可能是一小部分<=〜20)。
uint16_t
我有一个最简单的比较函数,可以想到使用unsigned long longs(您不能只减去它们):
unsigned long long
qsort(hpidx, num_pix, sizeof(uint64_t), comp_uint64); ... int comp_uint64(const void *a, const void *b) { if(*((uint64_t *)a) > *((uint64_t *)b)) return(+1); if(*((uint64_t *)a) < *((uint64_t *)b)) return(-1); return(0); } // End Comp_uint64().
StackExchange上有一个非常有趣的“ Programming Puzzles&Code Golf”,但他们使用floats。然后是QSort,RecQuick,堆,stooge,tree,radix …
float
swenson / sort看起来很有趣,但是没有(明显)支持我的数据类型uint64_t。而“快速排序”时间是最好的。一些消息来源说,系统qsort可以是任何东西,不一定是“快速排序”。
C ++排序绕过了void指针的通用转换,并实现了优于C的性能。必须有一种优化的方法,以64位处理器以翘曲速度猛击U8。
系统/编译器信息:
我目前正在将GCC与Strawberry Perl一起使用
gcc version 4.9.2 (x86_64-posix-sjlj, built by strawberryperl.com Intel 2700K Sandy Bridge CPU, 32GB DDR3 windows 7/64 pro gcc -D__USE_MINGW_ANSI_STDIO -O4 -ffast-math -m64 -Ofast -march=corei7-avx -mtune=corei7 -Ic:/bin/xxHash-master -Lc:/bin/xxHash-master c:/bin/stddev.c -o c:/bin/stddev.g6.exe
QSORT()
尝试使用Michael Tokarev的inline qsort。
“可以用了”?从qsort.h文档
qsort.h
----------------------------- * Several ready-to-use examples: * * Sorting array of integers: * void int_qsort(int *arr, unsigned n) { * #define int_lt(a,b) ((*a)<(*b)) * QSORT(int, arr, n, int_lt); -------------------------------- Change from type "int" to "uint64_t" compile error on TYPE??? c:/bin/bpbfct.c:586:8: error: expected expression before 'uint64_t' QSORT(uint64_t, hpidx, num_pix, islt);
我找不到真正的,可编译的,有效的示例程序,只是带有“一般概念”的注释
#define QSORT_TYPE uint64_t #define islt(a,b) ((*a)<(*b)) uint64_t *QSORT_BASE; int QSORT_NELT; hpidx=(uint64_t *) calloc(num_pix+2, sizeof(uint64_t)); // Hash . PIDX QSORT_BASE = hpidx; QSORT_NELT = num_pix; // QSORT_LT is function QSORT_LT() QSORT(uint64_t, hpidx, num_pix, islt); //QSORT(uint64_t *, hpidx, num_pix, QSORT_LT); // QSORT_LT mal-defined? //qsort(hpidx, num_pix, sizeof(uint64_t), comp_uint64); // << WORKS
“准备使用的”示例使用类型的int,char *和struct elt。不是uint64_t类型吗?尝试long long
int
char *
struct elt
long long
QSORT(long long, hpidx, num_pix, islt); c:/bin/bpbfct.c:586:8: error: expected expression before 'long' QSORT(long long, hpidx, num_pix, islt);
RADIXSORT
结果:RADIX_SORT是RADICAL!
I:\br3\pf.249465>grep "Event" bb12.log | grep -i Sort << 1.40 sec average 4) Time=1.411 sec = 49.61%, Event RADIX_SORT , hits=1 4) Time=1.396 sec = 49.13%, Event RADIX_SORT , hits=1 4) Time=1.392 sec = 49.15%, Event RADIX_SORT , hits=1 16) Time=1.414 sec = 49.12%, Event RADIX_SORT , hits=1 I:\br3\pf.249465>grep "Event" bb11.log | grep -i Sort << 5.525 sec average = 3.95 time slower 4) Time=5.538 sec = 86.34%, Event QSort , hits=1 4) Time=5.519 sec = 79.41%, Event QSort , hits=1 4) Time=5.519 sec = 79.02%, Event QSort , hits=1 4) Time=5.563 sec = 79.49%, Event QSort , hits=1 4) Time=5.684 sec = 79.83%, Event QSort , hits=1 4) Time=5.509 sec = 79.30%, Event QSort , hits=1
比qsort开箱即用的速度快3.94倍!
而且,更重要的是,有实际的,有效的代码,而不仅仅是一些大师提供的80%的代码,他们假设您知道他们所知道的一切,并且可以填写其他20%的代码。
很棒的解决方案!谢谢路易斯·里奇!
我将使用8位基数的Radix Sort。对于64位值,一个经过优化的基数排序将必须在列表上进行9次迭代(一次要预先计算计数和偏移,而64位/ 8位则需要8次)。9 * N时间和2 * N空间(使用阴影数组)。
这是优化的基数排序的样子。
typedef union { struct { uint32_t c8[256]; uint32_t c7[256]; uint32_t c6[256]; uint32_t c5[256]; uint32_t c4[256]; uint32_t c3[256]; uint32_t c2[256]; uint32_t c1[256]; }; uint32_t counts[256 * 8]; } rscounts_t; uint64_t * radixSort(uint64_t * array, uint32_t size) { rscounts_t counts; memset(&counts, 0, 256 * 8 * sizeof(uint32_t)); uint64_t * cpy = (uint64_t *)malloc(size * sizeof(uint64_t)); uint32_t o8=0, o7=0, o6=0, o5=0, o4=0, o3=0, o2=0, o1=0; uint32_t t8, t7, t6, t5, t4, t3, t2, t1; uint32_t x; // calculate counts for(x = 0; x < size; x++) { t8 = array[x] & 0xff; t7 = (array[x] >> 8) & 0xff; t6 = (array[x] >> 16) & 0xff; t5 = (array[x] >> 24) & 0xff; t4 = (array[x] >> 32) & 0xff; t3 = (array[x] >> 40) & 0xff; t2 = (array[x] >> 48) & 0xff; t1 = (array[x] >> 56) & 0xff; counts.c8[t8]++; counts.c7[t7]++; counts.c6[t6]++; counts.c5[t5]++; counts.c4[t4]++; counts.c3[t3]++; counts.c2[t2]++; counts.c1[t1]++; } // convert counts to offsets for(x = 0; x < 256; x++) { t8 = o8 + counts.c8[x]; t7 = o7 + counts.c7[x]; t6 = o6 + counts.c6[x]; t5 = o5 + counts.c5[x]; t4 = o4 + counts.c4[x]; t3 = o3 + counts.c3[x]; t2 = o2 + counts.c2[x]; t1 = o1 + counts.c1[x]; counts.c8[x] = o8; counts.c7[x] = o7; counts.c6[x] = o6; counts.c5[x] = o5; counts.c4[x] = o4; counts.c3[x] = o3; counts.c2[x] = o2; counts.c1[x] = o1; o8 = t8; o7 = t7; o6 = t6; o5 = t5; o4 = t4; o3 = t3; o2 = t2; o1 = t1; } // radix for(x = 0; x < size; x++) { t8 = array[x] & 0xff; cpy[counts.c8[t8]] = array[x]; counts.c8[t8]++; } for(x = 0; x < size; x++) { t7 = (cpy[x] >> 8) & 0xff; array[counts.c7[t7]] = cpy[x]; counts.c7[t7]++; } for(x = 0; x < size; x++) { t6 = (array[x] >> 16) & 0xff; cpy[counts.c6[t6]] = array[x]; counts.c6[t6]++; } for(x = 0; x < size; x++) { t5 = (cpy[x] >> 24) & 0xff; array[counts.c5[t5]] = cpy[x]; counts.c5[t5]++; } for(x = 0; x < size; x++) { t4 = (array[x] >> 32) & 0xff; cpy[counts.c4[t4]] = array[x]; counts.c4[t4]++; } for(x = 0; x < size; x++) { t3 = (cpy[x] >> 40) & 0xff; array[counts.c3[t3]] = cpy[x]; counts.c3[t3]++; } for(x = 0; x < size; x++) { t2 = (array[x] >> 48) & 0xff; cpy[counts.c2[t2]] = array[x]; counts.c2[t2]++; } for(x = 0; x < size; x++) { t1 = (cpy[x] >> 56) & 0xff; array[counts.c1[t1]] = cpy[x]; counts.c1[t1]++; } free(cpy); return array; }