问题
For code related to this question, I need to compute the following as fast as possible:
Given a 32 bit integer i, compute the position of the n-th least significant set bit. Both n and the result should be 0-indexed.
For example, given the number i = 110101101012 and n = 4, the desired number is 7 as the fourth set bit is at position 7: 11010110101.
Using the pdep instruction from the BMI2 instruction set extension for x86 and the commonly available __builtin_ctz() intrinsic function, this can be computed easily:
j = _pdep_u32(1 << n, i);
return (__builtin_ctz(j));
However, many computers do not have the pdep instruction (and it's slow on AMD CPUs that do have it), rendering this approach non-portable. You can also compute such bit positions without pdep like this:
j = i;
for (k = 0; k < n; k++)
j &= j - 1;
return (__builtin_ctz(j));
However, this is pretty slow.
I am targeting computers that provide at least __builtin_popcount() and __builtin_ctz(). How can I find such bit positions faster?
回答1:
The version from bit-twiddling hacks adapted to this case is, for example,
unsigned int nth_bit_set(uint32_t value, unsigned int n)
{
const uint32_t pop2 = (value & 0x55555555u) + ((value >> 1) & 0x55555555u);
const uint32_t pop4 = (pop2 & 0x33333333u) + ((pop2 >> 2) & 0x33333333u);
const uint32_t pop8 = (pop4 & 0x0f0f0f0fu) + ((pop4 >> 4) & 0x0f0f0f0fu);
const uint32_t pop16 = (pop8 & 0x00ff00ffu) + ((pop8 >> 8) & 0x00ff00ffu);
const uint32_t pop32 = (pop16 & 0x000000ffu) + ((pop16 >>16) & 0x000000ffu);
unsigned int rank = 0;
unsigned int temp;
if (n++ >= pop32)
return 32;
temp = pop16 & 0xffu;
/* if (n > temp) { n -= temp; rank += 16; } */
rank += ((temp - n) & 256) >> 4;
n -= temp & ((temp - n) >> 8);
temp = (pop8 >> rank) & 0xffu;
/* if (n > temp) { n -= temp; rank += 8; } */
rank += ((temp - n) & 256) >> 5;
n -= temp & ((temp - n) >> 8);
temp = (pop4 >> rank) & 0x0fu;
/* if (n > temp) { n -= temp; rank += 4; } */
rank += ((temp - n) & 256) >> 6;
n -= temp & ((temp - n) >> 8);
temp = (pop2 >> rank) & 0x03u;
/* if (n > temp) { n -= temp; rank += 2; } */
rank += ((temp - n) & 256) >> 7;
n -= temp & ((temp - n) >> 8);
temp = (value >> rank) & 0x01u;
/* if (n > temp) rank += 1; */
rank += ((temp - n) & 256) >> 8;
return rank;
}
which, when compiled in a separate compilation unit, on gcc-5.4.0 using -Wall -O3 -march=native -mtune=native on Intel Core i5-4200u, yields
00400a40 <nth_bit_set>:
400a40: 89 f9 mov %edi,%ecx
400a42: 89 f8 mov %edi,%eax
400a44: 55 push %rbp
400a45: 40 0f b6 f6 movzbl %sil,%esi
400a49: d1 e9 shr %ecx
400a4b: 25 55 55 55 55 and $0x55555555,%eax
400a50: 53 push %rbx
400a51: 81 e1 55 55 55 55 and $0x55555555,%ecx
400a57: 01 c1 add %eax,%ecx
400a59: 41 89 c8 mov %ecx,%r8d
400a5c: 89 c8 mov %ecx,%eax
400a5e: 41 c1 e8 02 shr $0x2,%r8d
400a62: 25 33 33 33 33 and $0x33333333,%eax
400a67: 41 81 e0 33 33 33 33 and $0x33333333,%r8d
400a6e: 41 01 c0 add %eax,%r8d
400a71: 45 89 c1 mov %r8d,%r9d
400a74: 44 89 c0 mov %r8d,%eax
400a77: 41 c1 e9 04 shr $0x4,%r9d
400a7b: 25 0f 0f 0f 0f and $0xf0f0f0f,%eax
400a80: 41 81 e1 0f 0f 0f 0f and $0xf0f0f0f,%r9d
400a87: 41 01 c1 add %eax,%r9d
400a8a: 44 89 c8 mov %r9d,%eax
400a8d: 44 89 ca mov %r9d,%edx
400a90: c1 e8 08 shr $0x8,%eax
400a93: 81 e2 ff 00 ff 00 and $0xff00ff,%edx
400a99: 25 ff 00 ff 00 and $0xff00ff,%eax
400a9e: 01 d0 add %edx,%eax
400aa0: 0f b6 d8 movzbl %al,%ebx
400aa3: c1 e8 10 shr $0x10,%eax
400aa6: 0f b6 d0 movzbl %al,%edx
400aa9: b8 20 00 00 00 mov $0x20,%eax
400aae: 01 da add %ebx,%edx
400ab0: 39 f2 cmp %esi,%edx
400ab2: 77 0c ja 400ac0 <nth_bit_set+0x80>
400ab4: 5b pop %rbx
400ab5: 5d pop %rbp
400ab6: c3 retq
400ac0: 83 c6 01 add $0x1,%esi
400ac3: 89 dd mov %ebx,%ebp
400ac5: 29 f5 sub %esi,%ebp
400ac7: 41 89 ea mov %ebp,%r10d
400aca: c1 ed 08 shr $0x8,%ebp
400acd: 41 81 e2 00 01 00 00 and $0x100,%r10d
400ad4: 21 eb and %ebp,%ebx
400ad6: 41 c1 ea 04 shr $0x4,%r10d
400ada: 29 de sub %ebx,%esi
400adc: c4 42 2b f7 c9 shrx %r10d,%r9d,%r9d
400ae1: 41 0f b6 d9 movzbl %r9b,%ebx
400ae5: 89 dd mov %ebx,%ebp
400ae7: 29 f5 sub %esi,%ebp
400ae9: 41 89 e9 mov %ebp,%r9d
400aec: 41 81 e1 00 01 00 00 and $0x100,%r9d
400af3: 41 c1 e9 05 shr $0x5,%r9d
400af7: 47 8d 14 11 lea (%r9,%r10,1),%r10d
400afb: 41 89 e9 mov %ebp,%r9d
400afe: 41 c1 e9 08 shr $0x8,%r9d
400b02: c4 42 2b f7 c0 shrx %r10d,%r8d,%r8d
400b07: 41 83 e0 0f and $0xf,%r8d
400b0b: 44 21 cb and %r9d,%ebx
400b0e: 45 89 c3 mov %r8d,%r11d
400b11: 29 de sub %ebx,%esi
400b13: 5b pop %rbx
400b14: 41 29 f3 sub %esi,%r11d
400b17: 5d pop %rbp
400b18: 44 89 da mov %r11d,%edx
400b1b: 41 c1 eb 08 shr $0x8,%r11d
400b1f: 81 e2 00 01 00 00 and $0x100,%edx
400b25: 45 21 d8 and %r11d,%r8d
400b28: c1 ea 06 shr $0x6,%edx
400b2b: 44 29 c6 sub %r8d,%esi
400b2e: 46 8d 0c 12 lea (%rdx,%r10,1),%r9d
400b32: c4 e2 33 f7 c9 shrx %r9d,%ecx,%ecx
400b37: 83 e1 03 and $0x3,%ecx
400b3a: 41 89 c8 mov %ecx,%r8d
400b3d: 41 29 f0 sub %esi,%r8d
400b40: 44 89 c0 mov %r8d,%eax
400b43: 41 c1 e8 08 shr $0x8,%r8d
400b47: 25 00 01 00 00 and $0x100,%eax
400b4c: 44 21 c1 and %r8d,%ecx
400b4f: c1 e8 07 shr $0x7,%eax
400b52: 29 ce sub %ecx,%esi
400b54: 42 8d 14 08 lea (%rax,%r9,1),%edx
400b58: c4 e2 6b f7 c7 shrx %edx,%edi,%eax
400b5d: 83 e0 01 and $0x1,%eax
400b60: 29 f0 sub %esi,%eax
400b62: 25 00 01 00 00 and $0x100,%eax
400b67: c1 e8 08 shr $0x8,%eax
400b6a: 01 d0 add %edx,%eax
400b6c: c3 retq
When compiled as a separate compilation unit, timing on this machine is difficult, because the actual operation is as fast as calling a do-nothing function (also compiled in a separate compilation unit); essentially, the calculation is done during the latencies associated with the function call.
It seems to be slightly faster than my suggestion of a binary search,
unsigned int nth_bit_set(uint32_t value, unsigned int n)
{
uint32_t mask = 0x0000FFFFu;
unsigned int size = 16u;
unsigned int base = 0u;
if (n++ >= __builtin_popcount(value))
return 32;
while (size > 0) {
const unsigned int count = __builtin_popcount(value & mask);
if (n > count) {
base += size;
size >>= 1;
mask |= mask << size;
} else {
size >>= 1;
mask >>= size;
}
}
return base;
}
where the loop is executed exactly five times, compiling to
00400ba0 <nth_bit_set>:
400ba0: 83 c6 01 add $0x1,%esi
400ba3: 31 c0 xor %eax,%eax
400ba5: b9 10 00 00 00 mov $0x10,%ecx
400baa: ba ff ff 00 00 mov $0xffff,%edx
400baf: 45 31 db xor %r11d,%r11d
400bb2: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
400bb8: 41 89 c9 mov %ecx,%r9d
400bbb: 41 89 f8 mov %edi,%r8d
400bbe: 41 d0 e9 shr %r9b
400bc1: 41 21 d0 and %edx,%r8d
400bc4: c4 62 31 f7 d2 shlx %r9d,%edx,%r10d
400bc9: f3 45 0f b8 c0 popcnt %r8d,%r8d
400bce: 41 09 d2 or %edx,%r10d
400bd1: 44 38 c6 cmp %r8b,%sil
400bd4: 41 0f 46 cb cmovbe %r11d,%ecx
400bd8: c4 e2 33 f7 d2 shrx %r9d,%edx,%edx
400bdd: 41 0f 47 d2 cmova %r10d,%edx
400be1: 01 c8 add %ecx,%eax
400be3: 44 89 c9 mov %r9d,%ecx
400be6: 45 84 c9 test %r9b,%r9b
400be9: 75 cd jne 400bb8 <nth_bit_set+0x18>
400beb: c3 retq
as in, not more than 31 cycles in 95% of calls to the binary search version, compared to not more than 28 cycles in 95% of calls to the bit-hack version; both run within 28 cycles in 50% of the cases. (The loop version takes up to 56 cycles in 95% of calls, up to 37 cycles median.)
To determine which one is better in actual real-world code, one would have to do a proper benchmark within the real-world task; at least with current x86-64 architecture processors, the work done is easily hidden in latencies incurred elsewhere (like function calls).
回答2:
My answer is mostly based on this implementation of a 64bit word select method (Hint: Look only at the MARISA_USE_POPCNT, MARISA_X64, MARISA_USE_SSE3 codepaths):
It works in two steps, first selecting the byte containing the n-th set bit and then using a lookup table inside the byte:
- Extract the lower and higher nibbles for every byte (bitmasks 0xF, 0xF0, shift the higher nibbles down)
- Replace the nibble values by their popcount (_mm_shuffle_epi8 with A000120)
- Sum the popcounts of the lower and upper nibbles (Normal SSE addition) to get byte popcounts
- Compute the prefix sum over all byte popcounts (multiplication with 0x01010101...)
- Propagate the position
nto all bytes (SSE broadcast or again multiplication with 0x01010101...) - Do a bytewise comparison (_mm_cmpgt_epi8 leaves 0xFF in every byte smaller than
n) - Compute the byte offset by doing a popcount on the result
Now we know which byte contains the bit and a simple byte lookup table like in grek40's answer suffices to get the result.
Note however that I have not really benchmarked this result against other implementations, only that I have seen it to be quite efficient (and branchless)
回答3:
Edit
After giving it some thought and using the __builtin_popcount function, I figured it might be better to decide on the relevant byte and then compute the whole result instead of incrementally adding/subtracting numbers. Here is an updated version:
int GetBitAtPosition(unsigned i, unsigned n)
{
unsigned bitCount;
bitCount = __builtin_popcount(i & 0x00ffffff);
if (bitCount <= n)
{
return (24 + LUT_BitPosition[i >> 24][n - bitCount]);
}
bitCount = __builtin_popcount(i & 0x0000ffff);
if (bitCount <= n)
{
return (16 + LUT_BitPosition[(i >> 16) & 0xff][n - bitCount]);
}
bitCount = __builtin_popcount(i & 0x000000ff);
if (bitCount <= n)
{
return (8 + LUT_BitPosition[(i >> 8) & 0xff][n - bitCount]);
}
return LUT_BitPosition[i & 0xff][n];
}
I felt like creating a LUT based solution where the number is inspected in byte-chunks, however, the LUT for the n-th bit position grew quite large (256*8) and the LUT-free version that was discussed in the comments might be better.
Generally the algorithm would look like this:
unsigned i = 0x000006B5;
unsigned n = 4;
unsigned result = 0;
unsigned bitCount;
while (i)
{
bitCount = LUT_BitCount[i & 0xff];
if (n < bitCount)
{
result += LUT_BitPosition[i & 0xff][n];
break; // found
}
else
{
n -= bitCount;
result += 8;
i >>= 8;
}
}
Might be worth to unroll the loop into its up to 4 iterations to get the best performance on 32 bit numbers.
The LUT for bitcount (could be replaced by __builtin_popcount):
unsigned LUT_BitCount[] = {
0, 1, 1, 2, 1, 2, 2, 3, // 0-7
1, 2, 2, 3, 2, 3, 3, 4, // 8-15
1, 2, 2, 3, 2, 3, 3, 4, // 16-23
2, 3, 3, 4, 3, 4, 4, 5, // 24-31
1, 2, 2, 3, 2, 3, 3, 4, // 32-39
2, 3, 3, 4, 3, 4, 4, 5, // 40-47
2, 3, 3, 4, 3, 4, 4, 5, // 48-55
3, 4, 4, 5, 4, 5, 5, 6, // 56-63
1, 2, 2, 3, 2, 3, 3, 4, // 64-71
2, 3, 3, 4, 3, 4, 4, 5, // 72-79
2, 3, 3, 4, 3, 4, 4, 5, // 80-87
3, 4, 4, 5, 4, 5, 5, 6, // 88-95
2, 3, 3, 4, 3, 4, 4, 5, // 96-103
3, 4, 4, 5, 4, 5, 5, 6, // 104-111
3, 4, 4, 5, 4, 5, 5, 6, // 112-119
4, 5, 5, 6, 5, 6, 6, 7, // 120-127
1, 2, 2, 3, 2, 3, 3, 4, // 128
2, 3, 3, 4, 3, 4, 4, 5, // 136
2, 3, 3, 4, 3, 4, 4, 5, // 144
3, 4, 4, 5, 4, 5, 5, 6, // 152
2, 3, 3, 4, 3, 4, 4, 5, // 160
3, 4, 4, 5, 4, 5, 5, 6, // 168
3, 4, 4, 5, 4, 5, 5, 6, // 176
4, 5, 5, 6, 5, 6, 6, 7, // 184
2, 3, 3, 4, 3, 4, 4, 5, // 192
3, 4, 4, 5, 4, 5, 5, 6, // 200
3, 4, 4, 5, 4, 5, 5, 6, // 208
4, 5, 5, 6, 5, 6, 6, 7, // 216
3, 4, 4, 5, 4, 5, 5, 6, // 224
4, 5, 5, 6, 5, 6, 6, 7, // 232
4, 5, 5, 6, 5, 6, 6, 7, // 240
5, 6, 6, 7, 6, 7, 7, 8, // 248-255
};
The LUT for bit position within a byte:
unsigned LUT_BitPosition[][8] = {
// 0-7
{UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
// 8-15
{3,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
// 16-31
{4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{3,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,4,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,4,UINT_MAX,UINT_MAX,UINT_MAX},
// 32-63
{5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{3,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,5,UINT_MAX,UINT_MAX,UINT_MAX},
{4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,4,5,UINT_MAX,UINT_MAX,UINT_MAX},
{3,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,4,5,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,4,5,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,4,5,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,4,5,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,4,5,UINT_MAX,UINT_MAX},
// 64-127
{6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{3,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,6,UINT_MAX,UINT_MAX,UINT_MAX},
{4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,4,6,UINT_MAX,UINT_MAX,UINT_MAX},
{3,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,4,6,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,4,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,4,6,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,4,6,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,4,6,UINT_MAX,UINT_MAX},
{5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{3,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,5,6,UINT_MAX,UINT_MAX},
{4,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{2,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,4,5,6,UINT_MAX,UINT_MAX},
{3,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,4,5,6,UINT_MAX,UINT_MAX},
{2,3,4,5,6,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,4,5,6,UINT_MAX,UINT_MAX},
{1,2,3,4,5,6,UINT_MAX,UINT_MAX},
{0,1,2,3,4,5,6,UINT_MAX},
// 128-255
{7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{3,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,7,UINT_MAX,UINT_MAX,UINT_MAX},
{4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,4,7,UINT_MAX,UINT_MAX,UINT_MAX},
{3,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,4,7,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,4,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,4,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,4,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,4,7,UINT_MAX,UINT_MAX},
{5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{3,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,5,7,UINT_MAX,UINT_MAX},
{4,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{2,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,4,5,7,UINT_MAX,UINT_MAX},
{3,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,4,5,7,UINT_MAX,UINT_MAX},
{2,3,4,5,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,4,5,7,UINT_MAX,UINT_MAX},
{1,2,3,4,5,7,UINT_MAX,UINT_MAX},
{0,1,2,3,4,5,7,UINT_MAX},
{6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{2,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{3,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{2,3,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,3,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,3,6,7,UINT_MAX,UINT_MAX},
{4,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{2,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,4,6,7,UINT_MAX,UINT_MAX},
{3,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,4,6,7,UINT_MAX,UINT_MAX},
{2,3,4,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,4,6,7,UINT_MAX,UINT_MAX},
{1,2,3,4,6,7,UINT_MAX,UINT_MAX},
{0,1,2,3,4,6,7,UINT_MAX},
{5,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{1,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{2,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,2,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,2,5,6,7,UINT_MAX,UINT_MAX},
{3,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,3,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,3,5,6,7,UINT_MAX,UINT_MAX},
{2,3,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,3,5,6,7,UINT_MAX,UINT_MAX},
{1,2,3,5,6,7,UINT_MAX,UINT_MAX},
{0,1,2,3,5,6,7,UINT_MAX},
{4,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX,UINT_MAX},
{0,4,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{1,4,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,1,4,5,6,7,UINT_MAX,UINT_MAX},
{2,4,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,2,4,5,6,7,UINT_MAX,UINT_MAX},
{1,2,4,5,6,7,UINT_MAX,UINT_MAX},
{0,1,2,4,5,6,7,UINT_MAX},
{3,4,5,6,7,UINT_MAX,UINT_MAX,UINT_MAX},
{0,3,4,5,6,7,UINT_MAX,UINT_MAX},
{1,3,4,5,6,7,UINT_MAX,UINT_MAX},
{0,1,3,4,5,6,7,UINT_MAX},
{2,3,4,5,6,7,UINT_MAX,UINT_MAX},
{0,2,3,4,5,6,7,UINT_MAX},
{1,2,3,4,5,6,7,UINT_MAX},
{0,1,2,3,4,5,6,7},
};
回答4:
My approach is to calculate the population count for each 8-bit quarters of the 32-bit integer in parallel, then find which quarter contains the nth bit. The population count of quarters that are lower than the found one can be summarized as the initial value of later calculation.
After that count set bits one-by-one until the n is reached. Without branches and using an incomplete implementation of population count algorithm, my example is the following:
#include <stdio.h>
#include <stdint.h>
int main() {
uint32_t n = 10, test = 3124375902u; /* 10111010001110100011000101011110 */
uint32_t index, popcnt, quarter = 0, q_popcnt;
/* count set bits of each quarter of 32-bit integer in parallel */
q_popcnt = test - ((test >> 1) & 0x55555555);
q_popcnt = (q_popcnt & 0x33333333) + ((q_popcnt >> 2) & 0x33333333);
q_popcnt = (q_popcnt + (q_popcnt >> 4)) & 0x0F0F0F0F;
popcnt = q_popcnt;
/* find which quarters can be summarized and summarize them */
quarter += (n + 1 >= (q_popcnt & 0xff));
quarter += (n + 1 >= ((q_popcnt += q_popcnt >> 8) & 0xff));
quarter += (n + 1 >= ((q_popcnt += q_popcnt >> 16) & 0xff));
quarter += (n + 1 >= ((q_popcnt += q_popcnt >> 24) & 0xff));
popcnt &= (UINT32_MAX >> (8 * quarter));
popcnt = (popcnt * 0x01010101) >> 24;
/* find the index of nth bit in quarter where it should be */
index = 8 * quarter;
index += ((popcnt += (test >> index) & 1) <= n);
index += ((popcnt += (test >> index) & 1) <= n);
index += ((popcnt += (test >> index) & 1) <= n);
index += ((popcnt += (test >> index) & 1) <= n);
index += ((popcnt += (test >> index) & 1) <= n);
index += ((popcnt += (test >> index) & 1) <= n);
index += ((popcnt += (test >> index) & 1) <= n);
index += ((popcnt += (test >> index) & 1) <= n);
printf("index = %u\n", index);
return 0;
}
A simple approach which uses loops and conditionals can be the following as well:
#include <stdio.h>
#include <stdint.h>
int main() {
uint32_t n = 11, test = 3124375902u; /* 10111010001110100011000101011110 */
uint32_t popcnt = 0, index = 0;
while(popcnt += ((test >> index) & 1), popcnt <= n && ++index < 32);
printf("index = %u\n", index);
return 0;
}
回答5:
Based on a method by Juha Järvi published in the famous Bit Twiddling Hacks, I tested this implementation where n and i are used as in the question:
a = i - (i >> 1 & 0x55555555);
b = (a & 0x33333333) + (a >> 2 & 0x33333333);
c = b + (b >> 4) & 0x0f0f0f0f;
r = n + 1;
s = 0;
t = c + (c >> 8) & 0xff;
if (r > t) {
s += 16;
r -= t;
}
t = c >> s & 0xf;
if (r > t) {
s += 8;
r -= t;
}
t = b >> s & 0x7;
if (r > t) {
s += 4;
r -= t;
}
t = a >> s & 0x3;
if (r > t) {
s += 2;
r -= t;
}
t = i >> s & 0x1;
if (r > t)
s++;
return (s);
Based on my own tests, this is about as fast as the loop on x86, whereas it is 20% faster on arm64 and probably a lot faster on arm due to the fast conditional instructions, but I can't test this right now.
来源：https://stackoverflow.com/questions/45482787/how-to-efficiently-find-the-n-th-set-bit