BeeOnRope BeeOnRope - 1 year ago 60
C Question

Efficiently dividing unsigned value by a power of two, rounding up

I want to implement unsigned integer division by an arbitrary power of two, rounding up, efficiently. So what I, mathematically, is

0. In C, the straw-man implementation, which doesn't take advantage of the restricted domain of
would something like the following function1:

/** q must be a power of 2 */
uint64_t divide(uint64_t p, uint64_t q) {
uint64_t res = p / q;
return p % q == 0 ? res : res + 1;

... of course, I don't actually want to use division or mod at the machine level, since that takes many cycles even on modern hardware. I'm looking for a strength reduction that uses shifts and/or some other cheap operation(s) - taking advantage of the fact that
is a power of 2.

You can assume we have an efficient
lg(unsigned int x)
function, which returns the base-2 log of
, if
is a power-of-two.

Undefined behavior is fine if
is zero.

Please note that the simple solution:
(p+q-1) >> lg(q)
doesn't work in general - try it with
p == 2^64-100 and q == 256
2 for example.

Platform Details

I'm interested in solutions in C, that are likely to perform well across a variety of platforms, but for the sake of concreteness, awarding the bounty and because any definitive discussion of performance needs to include a target architecture, I'll be specific about how I'll test them:

  • Skylake CPU

  • gcc 5.4.0
    with compile flags
    -O3 -march=haswell

Using gcc builtins (such as bitscan/leading zero builtins) is fine, and in particular I've implemented the
function I said was available as follows:

inline uint64_t lg(uint64_t x) {
return 63U - (uint64_t)__builtin_clzl(x);

inline uint32_t lg32(uint32_t x) {
return 31U - (uint32_t)__builtin_clz(x);

I verified that these compile down to a single
instruction, at least with
, despite the apparent involvement of a subtraction. You are of course free to ignore these and use whatever other builtins you want in your solution.


I wrote a benchmark for the existing answers, and will share and update the results as changes are made.

Writing a good benchmark for a small, potentially inlined operation is quite tough. When code is inlined into a call site, a lot of the work of the function may disappear, especially when it's in a loop3.

You could simply avoid the whole inlining problem by ensuring your code isn't inlined: declare it in another compilation unit. I tried to that with the
binary, but really the results are fairly pointless. Nearly all implementations tied at 4 or 5 cycles per call, but even a dummy method that does nothing other than
return 0
takes the same time. So you are mostly just measuring the
call + ret
overhead. Furthermore, you are almost never really going to use the functions like this - unless you messed up, they'll be available for inlining and that changes everything.

So the two benchmarks I'll focus the most on repeatedly call the method under test in a loop, allowing inlining, cross-function optmization, loop hoisting and even vectorization.

There are two overall benchmark types: latency and throughput. The key difference is that in the latency benchmark, each call to
is dependent on the previous call, so in general calls cannot be easily overlapped4:

uint32_t bench_divide_latency(uint32_t p, uint32_t q) {
uint32_t total = p;
for (unsigned i=0; i < ITERS; i++) {
total += divide_algo(total, q);
q = rotl1(q);
return total;

Note that the running
depends so on the output of each divide call, and that it is also an input to the

The throughput variant, on the other hand, doesn't feed the output of one divide into the subsequent one. This allows work from one call to be overlapped with a subsequent one (both by the compiler, but especially the CPU), and even allows vectorization:

uint32_t bench_divide_throughput(uint32_t p, uint32_t q) {
uint32_t total = p;
for (unsigned i=0; i < ITERS; i++) {
total += fname(i, q);
q = rotl1(q);
return total;

Note that here we feed in the loop counter as the the dividend - this is variable, but it doesn't depend on the previous divide call.

Furthermore, each benchmark has three flavors of behavior for the divisor,

  1. Compile-time constant divisor. For example, a call to
    divide(p, 8)
    . This is common in practice, and the code can be much simpler when the divisor is known at compile time.

  2. Invariant divisor. Here the divisor is not know at compile time, but is constant for the whole benchmarking loop. This allows a subset of the optimizations that the compile-time constant does.

  3. Variable divisor. The divisor changes on each iteration of the loop. The benchmark functions above show this variant, using a "rotate left 1" instruction to vary the divisor.

Combining everything you get a total of 6 distinct benchmarks.



The results below test each algorithm over 1e9 iterations. Cycles are calculated simply by multiplying the time/call by the clock frequency. You can generally assume that something like
is the same as
, but the larger deviations like
seem to be real and reproducible.

The results for
(p + q - 1) >> lg(q)
but are only shown for reference, since this function fails for large
p + q
. The results for
are a very simple function:
return p + q
, and lets you estimate the benchmark overhead5 (the addition itself should take a cycle at most).

Bench: Compile-time constant Q
Function ns/call cycles
stoke32_32 1.93 5.00
divide_chux_32 1.55 4.01
divide_chux_64 1.55 4.01
divide_user23_32 1.97 5.11
divide_user23_64 1.93 5.00
divide_user23_variant_32 1.55 4.01
divide_user23_variant_64 1.55 4.01
divide_chrisdodd_32 1.95 5.05
divide_chrisdodd_64 1.93 4.99
divide_chris_32 4.63 12.00
divide_chris_64 4.52 11.71
divide_weather_32 2.72 7.04
divide_weather_64 2.78 7.20
divide_plusq_32 1.16 3.00
divide_plusq_64 1.16 3.00
divide_dummy_32 1.16 3.00
divide_dummy_64 1.16 3.00

Bench: Invariant Q
Function ns/call cycles
stoke32_32 1.93 5.00
divide_chux_32 1.56 4.03
divide_chux_64 1.55 4.01
divide_user23_32 1.95 5.04
divide_user23_64 1.99 5.17
divide_user23_variant_32 1.56 4.04
divide_user23_variant_64 1.55 4.01
divide_chrisdodd_32 1.95 5.04
divide_chrisdodd_64 1.93 4.99
divide_chris_32 4.60 11.90
divide_chris_64 4.57 11.85
divide_weather_32 12.54 32.48
divide_weather_64 19.18 49.69
divide_plusq_32 1.16 3.00
divide_plusq_64 1.16 3.00
divide_dummy_32 0.39 1.00
divide_dummy_64 0.39 1.00

Bench: Variable Q
Function ns/call cycles
stoke32_32 2.06 5.33
divide_chux_32 2.04 5.28
divide_chux_64 2.04 5.30
divide_user23_32 2.04 5.28
divide_user23_64 2.06 5.33
divide_user23_variant_32 2.04 5.27
divide_user23_variant_64 2.15 5.58
divide_chrisdodd_32 2.04 5.29
divide_chrisdodd_64 2.05 5.31
divide_chris_32 4.64 12.03
divide_chris_64 4.64 12.01
divide_weather_32 12.54 32.49
divide_weather_64 17.40 45.07
divide_plusq_32 1.93 4.99
divide_plusq_64 1.99 5.16
divide_dummy_32 0.39 1.02
divide_dummy_64 0.40 1.05


Here are the results for the throughput tests. Note that many of the algorithms here were auto-vectorized, so the performance is relatively very good for those: a fraction of a cycle in many cases. One result is that unlike most latency results, the 64-bit functions are considerably slower, since vectorization is more effective with smaller element sizes (although the gap is larger that I would have expected).

Bench: Compile-time constant Q
Function ns/call cycles
stoke32_32 0.39 1.00
divide_chux_32 0.15 0.39
divide_chux_64 0.53 1.37
divide_user23_32 0.14 0.36
divide_user23_64 0.53 1.37
divide_user23_variant_32 0.15 0.39
divide_user23_variant_64 0.53 1.37
divide_chrisdodd_32 1.16 3.00
divide_chrisdodd_64 1.16 3.00
divide_chris_32 4.34 11.23
divide_chris_64 4.34 11.24
divide_weather_32 1.35 3.50
divide_weather_64 1.35 3.50
divide_plusq_32 0.10 0.26
divide_plusq_64 0.39 1.00
divide_dummy_32 0.08 0.20
divide_dummy_64 0.39 1.00

Bench: Invariant Q
Function ns/call cycles
stoke32_32 0.48 1.25
divide_chux_32 0.15 0.39
divide_chux_64 0.48 1.25
divide_user23_32 0.17 0.43
divide_user23_64 0.58 1.50
divide_user23_variant_32 0.15 0.38
divide_user23_variant_64 0.48 1.25
divide_chrisdodd_32 1.16 3.00
divide_chrisdodd_64 1.16 3.00
divide_chris_32 4.35 11.26
divide_chris_64 4.36 11.28
divide_weather_32 5.79 14.99
divide_weather_64 17.00 44.02
divide_plusq_32 0.12 0.31
divide_plusq_64 0.48 1.25
divide_dummy_32 0.09 0.23
divide_dummy_64 0.09 0.23

Bench: Variable Q
Function ns/call cycles
stoke32_32 1.16 3.00
divide_chux_32 1.36 3.51
divide_chux_64 1.35 3.50
divide_user23_32 1.54 4.00
divide_user23_64 1.54 4.00
divide_user23_variant_32 1.36 3.51
divide_user23_variant_64 1.55 4.01
divide_chrisdodd_32 1.16 3.00
divide_chrisdodd_64 1.16 3.00
divide_chris_32 4.02 10.41
divide_chris_64 3.84 9.95
divide_weather_32 5.40 13.98
divide_weather_64 19.04 49.30
divide_plusq_32 1.03 2.66
divide_plusq_64 1.03 2.68
divide_dummy_32 0.63 1.63
divide_dummy_64 0.66 1.71

0 Of course, this notation doesn't actually work in C where
truncates the result so the
does nothing. So consider that pseudo-notation rather than straight C.

1 I'm also interested solutions where all types are
rather than

2 In general, any
p + q >= 2^64
causes an issue, due to overflow.

3 That said, the function should be in a loop, because the performance of a microscopic function that takes half a dozen cycles only really matters if it is called in a fairly tight loop.

4 This is a bit of a simplification - only the dividend
is dependent on the output of the previous iteration, so some work related to processing of
can still be overlapped.

5 Use such estimates with caution however - overhead isn't simply additive. If the overhead shows up as 4 cycles and some function
takes 5, it's likely not accurate to say the cost of the real work in
5 - 4 == 1
, because of the way execution is overlapped.

Answer Source

This answer is about what's ideal in asm; what we'd like to convince the compiler to emit for us. (I'm not suggesting actually using inline asm, except as a point of comparison when benchmarking compiler output.

I did manage to get pretty good asm output from pure C for ceil_div_andmask, see below.

It looks like Chris Dodd's general idea of return ((p-1) >> lg(q)) + 1 with special-case handling for d=0 is one of the best options. I.e. the optimal implementation of it in asm is hard to beat with an optimal implementation of anything else. Chux's (p >> lg(q)) + (bool)(p & (q-1)) also has advantages (like lower latency from p->result), and more CSE when the same q is used for multiple divisions. See below for a latency/throughput analysis of what gcc does with it.

If the same e = lg(q) is reused for multiple dividends, or the same dividend is reused for multiple divisors, different implementations can CSE more of the expression. They can also effectively vectorize with AVX2.

Branches are cheap and very efficient if they predict very well, so branching on d==0 will be best if it's almost never taken. If d==0 is not rare, branchless asm will perform better on average. Ideally we can write something in C that will let gcc make the right choice during profile-guided optimization, and compiles to good asm for either case.

Since the best branchless asm implementations don't add much latency I can see are (e.g. happens more than 1% or 5% of the time (or less than 95% of the time))

It's hard to guide gcc5.4 into emitting anything optimal. This is my work-in-progress on Godbolt).

I think the optimal sequences for Skylake for this algorithm are as follows. (Shown as stand-alone functions for the AMD64 SysV ABI, but talking about throughput/latency on the assumption that the compiler will emit something similar inlined into a loop, with no RET attached).

Branch on carry from calculating d-1 to detect d==0, instead of a separate test & branch. Reduces the uop count nicely, esp on SnB-family where JC can macro-fuse with SUB.

 xor    eax,eax          ; can take this uop off the fast path by adding a separate xor-and-return block, but in reality we want to inline something like this.
 sub    rdi, 1
 jc    .d_was_zero       ; fuses with the sub on SnB-family
 tzcnt  rax, rsi         ; tzcnt rsi,rsi also avoids any false-dep problems, but this illustrates that the q input can be read-only.
 shrx   rax, rdi, rax
 inc    rax
  • Fused-domain uops: 5 (not counting ret), and one of them is an xor-zero (no execution unit)
  • HSW/SKL latency with successful branch prediction:
    • (d==0): No data dependency on d or q, breaks the dep chain. (control dependency on d to detect mispredicts and retire the branch).
    • (d!=0): q->result: tzcnt+shrx+inc = 5c
    • (d!=0): d->result: sub+shrx+inc = 3c
  • Throughput: probably just bottlenecked on uop throughput

I've tried but failed to get gcc to branch on CF from the subtract, but it always wants to do a separate comparison. I know gcc can be coaxed into branching on CF after subtracting two variables, but maybe this fails if one is a compile-time constant. (IIRC, this typically compiles to a CF test with unsigned vars: foo -= bar; if(foo>bar) carry_detected = 1;)

Branchless with ADC / SBB to handle the d==0 case. Zero-handling adds only one instruction to the critical path (vs. a version with no special handling for d==0), but also converts one other from an INC to a sbb rax, -1 to make CF undo the -= -1. Using a CMOV is cheaper on pre-Broadwell, but takes extra instructions to set it up.

 tzcnt  rsi, rsi
 sub    rdi, 1
 adc    rdi, 0          ; d? d-1 : d.  Sets CF=CF
 shrx   rax, rdi, rsi
 sbb    rax, -1         ; result++ if d was non-zero
  • Fused-domain uops: 5 (not counting ret) on SKL. 7 on HSW
  • SKL latency:
    • q->result: tzcnt+shrx+sbb = 5c
    • d->result: sub+adc+shrx(dep on q begins here)+sbb = 4c
  • Throughput: TZCNT runs on p1. SBB, ADC, and SHRX only run on p06. So I think we bottleneck on 3 uops for p06 per iteration, making this run at best one iteration per 1.5c.

If q and d become ready at the same time, note that this version can run SUB/ADC in parallel with the 3c latency of TZCNT. If both are coming from the same cache-miss cache line, it's certainly possible. In any case, introducing the dep on q as late as possible in the d->result dependency chain is an advantage.

Getting this from C seems unlikely with gcc5.4. There is an intrinsic for add-with-carry, but gcc makes a total mess of it. It doesn't use immediate operands for ADC or SBB, and stores the carry into an integer reg between every operation. gcc7, clang3.9, and icc17 all make terrible code from this.

#include <x86intrin.h>
// compiles to completely horrible code, putting the flags into integer regs between ops.
T ceil_div_adc(T d, T q) {
  T e = lg(q);
  unsigned long long dm1;  // unsigned __int64
  unsigned char CF = _addcarry_u64(0, d, -1, &dm1);
  CF = _addcarry_u64(CF, 0, dm1, &dm1);
  T shifted = dm1 >> e;
  _subborrow_u64(CF, shifted, -1, &dm1);
  return dm1;
    # gcc5.4 -O3 -march=haswell
    mov     rax, -1
    tzcnt   rsi, rsi
    add     rdi, rax
    setc    cl
    xor     edx, edx
    add     cl, -1
    adc     rdi, rdx
    setc    dl
    shrx    rdi, rdi, rsi
    add     dl, -1
    sbb     rax, rdi

CMOV to fix the whole result: worse latency from q->result, since it's used sooner in the d->result dep chain.

 tzcnt  rsi, rsi
 sub    rdi, 1
 shrx   rax, rdi, rsi
 lea    rax, [rax+1]     ; inc preserving flags
 cmovc  rax, zeroed_register
  • Fused-domain uops: 5 (not counting ret) on SKL. 6 on HSW
  • SKL latency:
    • q->result: tzcnt+shrx+lea+cmov = 6c (worse than ADC/SBB by 1c)
    • d->result: sub+shrx(dep on q begins here)+lea+cmov = 4c
  • Throughput: TZCNT runs on p1. LEA is p15. CMOV and SHRX are p06. SUB is p0156. In theory only bottlenecked on fused-domain uop throughput, so one iteration per 1.25c. With lots of independent operations, resource conflicts from SUB or LEA stealing p1 or p06 shouldn't be a throughput problem because at 1 iter per 1.25c, no port is saturated with uops that can only run on that port.

CMOV to get an operand for SUB: I was hoping I could find a way to create an operand for a later instruction that would produce a zero when needed, without an input dependency on q, e, or the SHRX result. This would help if d is ready before q, or at the same time.

This doesn't achieve that goal, and needs an extra 7-byte mov rdx,-1 in the loop.

 tzcnt  rsi, rsi
 mov    rdx, -1
 sub    rdi, 1
 shrx   rax, rdi, rsi
 cmovnc rdx, rax
 sub    rax, rdx       ; res += d ? 1 : -res

Lower-latency version for pre-BDW CPUs with expensive CMOV, using SETCC to create a mask for AND.

 xor    edx, edx        ; needed every iteration

 tzcnt  rsi, rsi
 sub    rdi, 1

 setc   dl              ; d!=0 ?  0 : 1
 dec    rdx             ; d!=0 ? -1 : 0   // AND-mask

 shrx   rax, rdi, rsi
 inc    rax
 and    rax, rdx        ; zero the bogus result if d was initially 0

Still 4c latency from d->result (and 6 from q->result), because the SETC/DEC happen in parallel with the SHRX/INC. Total uop count: 8. Most of these insns can run on any port, so it should be 1 iter per 2 clocks.

Of course, for pre-HSW, you also need to replace SHRX.

We can get gcc5.4 to emit something nearly as good: (still uses a separate TEST instead of setting mask based on sub rdi, 1, but otherwise the same instructions as above). See it on Godbolt.

T ceil_div_andmask(T p, T q) {
    T mask = -(T)(p!=0);  // TEST+SETCC+NEG
    T e = lg(q);
    T nonzero_result = ((p-1) >> e) + 1;
    return nonzero_result & mask;

When the compiler knows that p is non-zero, it takes advantage and makes nice code:

#if defined(__GNUC__) && !defined(__INTEL_COMPILER)
#define assume(x) do{if(!(x)) __builtin_unreachable();}while(0)
#define assume(x) (void)(x) // still evaluate it once, for side effects in case anyone is insane enough to put any inside an assume()

T ceil_div_andmask_nonzerop(T p, T q) {
  return ceil_div_andmask(p, q);
    # gcc5.4 -O3 -march=haswell
    xor     eax, eax             # gcc7 does tzcnt in-place instead of wasting an insn on this
    sub     rdi, 1
    tzcnt   rax, rsi
    shrx    rax, rdi, rax
    add     rax, 1

Chux / user23_variant

only 3c latency from p->result, and constant q can CSE a lot.

T divide_A_chux(T p, T q) {
  bool round_up = p & (q-1);  // compiles differently from user23_variant with clang: AND instead of 
  return (p >> lg(q)) + round_up;

    xor     eax, eax           # in-place tzcnt would save this
    xor     edx, edx           # target for setcc
    tzcnt   rax, rsi
    sub     rsi, 1
    test    rsi, rdi
    shrx    rdi, rdi, rax
    setne   dl
    lea     rax, [rdx+rdi]

Doing the SETCC before TZCNT would allow an in-place TZCNT, saving the xor eax,eax. I haven't looked at how this inlines in a loop.

  • Fused-domain uops: 8 (not counting ret) on HSW/SKL
  • HSW/SKL latency:
    • q->result: (tzcnt+shrx(p) | sub+test(p)+setne) + lea(or add) = 5c
    • d->result: test(dep on q begins here)+setne+lea = 3c. (the shrx->lea chain is shorter, and thus not the critical path)
  • Throughput: Probably just bottlenecked on the frontend, at one iter per 2c. Saving the xor eax,eax should speed this up to one per 1.75c (but of course any loop overhead will be part of the bottleneck, because frontend bottlenecks are like that).
Recommended from our users: Dynamic Network Monitoring from WhatsUp Gold from IPSwitch. Free Download