Tuesday, March 26, 2019

performance - C++ code for testing the Collatz conjecture faster than hand-written assembly - why?



I wrote these two solutions for Project Euler Q14, in assembly and in C++. They are the same identical brute force approach for testing the Collatz conjecture. The assembly solution was assembled with



nasm -felf64 p14.asm && gcc p14.o -o p14



The C++ was compiled with



g++ p14.cpp -o p14


Assembly, p14.asm



section .data
fmt db "%d", 10, 0


global main
extern printf

section .text

main:
mov rcx, 1000000
xor rdi, rdi ; max i
xor rsi, rsi ; i


l1:
dec rcx
xor r10, r10 ; count
mov rax, rcx

l2:
test rax, 1
jpe even


mov rbx, 3
mul rbx
inc rax
jmp c1

even:
mov rbx, 2
xor rdx, rdx
div rbx


c1:
inc r10
cmp rax, 1
jne l2

cmp rdi, r10
cmovl rdi, r10
cmovl rsi, rcx

cmp rcx, 2

jne l1

mov rdi, fmt
xor rax, rax
call printf
ret


C++, p14.cpp




#include 

using namespace std;

int sequence(long n) {
int count = 1;
while (n != 1) {
if (n % 2 == 0)
n /= 2;
else

n = n*3 + 1;

++count;
}

return count;
}

int main() {
int max = 0, maxi;

for (int i = 999999; i > 0; --i) {
int s = sequence(i);
if (s > max) {
max = s;
maxi = i;
}
}

cout << maxi << endl;
}



I know about the compiler optimizations to improve speed and everything, but I don't see many ways to optimize my assembly solution further (speaking programmatically not mathematically).



The C++ code has modulus every term and division every even term, where assembly is only one division per even term.



But the assembly is taking on average 1 second longer than the C++ solution. Why is this? I am asking out of mainly curiosity.



Execution times




My system: 64 bit Linux on ‎1.4 GHz Intel Celeron 2955U (Haswell microarchitecture).




Answer



If you think a 64-bit DIV instruction is a good way to divide by two, then no wonder the compiler's asm output beat your hand-written code, even with -O0 (compile fast, no extra optimization, and store/reload to memory after/before every C statement so a debugger can modify variables).



See Agner Fog's Optimizing Assembly guide to learn how to write efficient asm. He also has instruction tables and a microarch guide for specific details for specific CPUs. See also the tag wiki for more perf links.



See also this more general question about beating the compiler with hand-written asm: Is inline assembly language slower than native C++ code?. TL:DR: yes if you do it wrong (like this question).




Usually you're fine letting the compiler do its thing, especially if you try to write C++ that can compile efficiently. Also see is assembly faster than compiled languages?. One of the answers links to these neat slides showing how various C compilers optimize some really simple functions with cool tricks.






even:
mov rbx, 2
xor rdx, rdx
div rbx



On Intel Haswell, div r64 is 36 uops, with a latency of 32-96 cycles, and a throughput of one per 21-74 cycles. (Plus the 2 uops to set up RBX and zero RDX, but out-of-order execution can run those early). High-uop-count instructions like DIV are microcoded, which can also cause front-end bottlenecks. In this case, latency is the most relevant factor because it's part of a loop-carried dependency chain.



shr rax, 1 does the same unsigned division: It's 1 uop, with 1c latency, and can run 2 per clock cycle.



For comparison, 32-bit division is faster, but still horrible vs. shifts. idiv r32 is 9 uops, 22-29c latency, and one per 8-11c throughput on Haswell.






As you can see from looking at gcc's -O0 asm output (Godbolt compiler explorer), it only uses shifts instructions. clang -O0 does compile naively like you thought, even using 64-bit IDIV twice. (When optimizing, compilers do use both outputs of IDIV when the source does a division and modulus with the same operands, if they use IDIV at all)




GCC doesn't have a totally-naive mode; it always transforms through GIMPLE, which means some "optimizations" can't be disabled. This includes recognizing division-by-constant and using shifts (power of 2) or a fixed-point multiplicative inverse (non power of 2) to avoid IDIV (see div_by_13 in the above godbolt link).



gcc -Os (optimize for size) does use IDIV for non-power-of-2 division,
unfortunately even in cases where the multiplicative inverse code is only slightly larger but much faster.








(summary for this case: use uint64_t n)




First of all, it's only interesting to look at optimized compiler output. (-O3). -O0 speed is basically meaningless.



Look at your asm output (on Godbolt, or see How to remove "noise" from GCC/clang assembly output?). When the compiler doesn't make optimal code in the first place: Writing your C/C++ source in a way that guides the compiler into making better code is usually the best approach. You have to know asm, and know what's efficient, but you apply this knowledge indirectly. Compilers are also a good source of ideas: sometimes clang will do something cool, and you can hand-hold gcc into doing the same thing: see this answer and what I did with the non-unrolled loop in @Veedrac's code below.)



This approach is portable, and in 20 years some future compiler can compile it to whatever is efficient on future hardware (x86 or not), maybe using new ISA extension or auto-vectorizing. Hand-written x86-64 asm from 15 years ago would usually not be optimally tuned for Skylake. e.g. compare&branch macro-fusion didn't exist back then. What's optimal now for hand-crafted asm for one microarchitecture might not be optimal for other current and future CPUs. Comments on @johnfound's answer discuss major differences between AMD Bulldozer and Intel Haswell, which have a big effect on this code. But in theory, g++ -O3 -march=bdver3 and g++ -O3 -march=skylake will do the right thing. (Or -march=native.) Or -mtune=... to just tune, without using instructions that other CPUs might not support.



My feeling is that guiding the compiler to asm that's good for a current CPU you care about shouldn't be a problem for future compilers. They're hopefully better than current compilers at finding ways to transform code, and can find a way that works for future CPUs. Regardless, future x86 probably won't be terrible at anything that's good on current x86, and the future compiler will avoid any asm-specific pitfalls while implementing something like the data movement from your C source, if it doesn't see something better.



Hand-written asm is a black-box for the optimizer, so constant-propagation doesn't work when inlining makes an input a compile-time constant. Other optimizations are also affected. Read https://gcc.gnu.org/wiki/DontUseInlineAsm before using asm. (And avoid MSVC-style inline asm: inputs/outputs have to go through memory which adds overhead.)




In this case: your n has a signed type, and gcc uses the SAR/SHR/ADD sequence that gives the correct rounding. (IDIV and arithmetic-shift "round" differently for negative inputs, see the SAR insn set ref manual entry). (IDK if gcc tried and failed to prove that n can't be negative, or what. Signed-overflow is undefined behaviour, so it should have been able to.)



You should have used uint64_t n, so it can just SHR. And so it's portable to systems where long is only 32-bit (e.g. x86-64 Windows).






BTW, gcc's optimized asm output looks pretty good (using unsigned long n): the inner loop it inlines into main() does this:



 # from gcc5.4 -O3  plus my comments


# edx= count=1
# rax= uint64_t n

.L9: # do{
lea rcx, [rax+1+rax*2] # rcx = 3*n + 1
mov rdi, rax
shr rdi # rdi = n>>1;
test al, 1 # set flags based on n%2 (aka n&1)
mov rax, rcx

cmove rax, rdi # n= (n%2) ? 3*n+1 : n/2;
add edx, 1 # ++count;
cmp rax, 1
jne .L9 #}while(n!=1)

cmp/branch to update max and maxi, and then do the next n


The inner loop is branchless, and the critical path of the loop-carried dependency chain is:





  • 3-component LEA (3 cycles)

  • cmov (2 cycles on Haswell, 1c on Broadwell or later).



Total: 5 cycle per iteration, latency bottleneck. Out-of-order execution takes care of everything else in parallel with this (in theory: I haven't tested with perf counters to see if it really runs at 5c/iter).



The FLAGS input of cmov (produced by TEST) is faster to produce than the RAX input (from LEA->MOV), so it's not on the critical path.



Similarly, the MOV->SHR that produces CMOV's RDI input is off the critical path, because it's also faster than the LEA. MOV on IvyBridge and later has zero latency (handled at register-rename time). (It still takes a uop, and a slot in the pipeline, so it's not free, just zero latency). The extra MOV in the LEA dep chain is part of the bottleneck on other CPUs.




The cmp/jne is also not part of the critical path: it's not loop-carried, because control dependencies are handled with branch prediction + speculative execution, unlike data dependencies on the critical path.








GCC did a pretty good job here. It could save one code byte by using inc edx instead of add edx, 1, because nobody cares about P4 and its false-dependencies for partial-flag-modifying instructions.



It could also save all the MOV instructions, and the TEST: SHR sets CF= the bit shifted out, so we can use cmovc instead of test / cmovz.




 ### Hand-optimized version of what gcc does
.L9: #do{
lea rcx, [rax+1+rax*2] # rcx = 3*n + 1
shr rax, 1 # n>>=1; CF = n&1 = n%2
cmovc rax, rcx # n= (n&1) ? 3*n+1 : n/2;
inc edx # ++count;
cmp rax, 1
jne .L9 #}while(n!=1)



See @johnfound's answer for another clever trick: remove the CMP by branching on SHR's flag result as well as using it for CMOV: zero only if n was 1 (or 0) to start with. (Fun fact: SHR with count != 1 on Nehalem or earlier causes a stall if you read the flag results. That's how they made it single-uop. The shift-by-1 special encoding is fine, though.)



Avoiding MOV doesn't help with the latency at all on Haswell (Can x86's MOV really be "free"? Why can't I reproduce this at all?). It does help significantly on CPUs like Intel pre-IvB, and AMD Bulldozer-family, where MOV is not zero-latency. The compiler's wasted MOV instructions do affect the critical path. BD's complex-LEA and CMOV are both lower latency (2c and 1c respectively), so it's a bigger fraction of the latency. Also, throughput bottlenecks become an issue, because it only has two integer ALU pipes. See @johnfound's answer, where he has timing results from an AMD CPU.



Even on Haswell, this version may help a bit by avoiding some occasional delays where a non-critical uop steals an execution port from one on the critical path, delaying execution by 1 cycle. (This is called a resource conflict). It also saves a register, which may help when doing multiple n values in parallel in an interleaved loop (see below).



LEA's latency depends on the addressing mode, on Intel SnB-family CPUs. 3c for 3 components ([base+idx+const], which takes two separate adds), but only 1c with 2 or fewer components (one add). Some CPUs (like Core2) do even a 3-component LEA in a single cycle, but SnB-family doesn't. Worse, Intel SnB-family standardizes latencies so there are no 2c uops, otherwise 3-component LEA would be only 2c like Bulldozer. (3-component LEA is slower on AMD as well, just not by as much).



So lea rcx, [rax + rax*2] / inc rcx is only 2c latency, faster than lea rcx, [rax + rax*2 + 1], on Intel SnB-family CPUs like Haswell. Break-even on BD, and worse on Core2. It does cost an extra uop, which normally isn't worth it to save 1c latency, but latency is the major bottleneck here and Haswell has a wide enough pipeline to handle the extra uop throughput.




Neither gcc, icc, nor clang (on godbolt) used SHR's CF output, always using an AND or TEST. Silly compilers. :P They're great pieces of complex machinery, but a clever human can often beat them on small-scale problems. (Given thousands to millions of times longer to think about it, of course! Compilers don't use exhaustive algorithms to search for every possible way to do things, because that would take too long when optimizing a lot of inlined code, which is what they do best. They also don't model the pipeline in the target microarchitecture, at least not in the same detail as IACA or other static-analysis tools; they just use some heuristics.)






Simple loop unrolling won't help; this loop bottlenecks on the latency of a loop-carried dependency chain, not on loop overhead / throughput. This means it would do well with hyperthreading (or any other kind of SMT), since the CPU has lots of time to interleave instructions from two threads. This would mean parallelizing the loop in main, but that's fine because each thread can just check a range of n values and produce a pair of integers as a result.



Interleaving by hand within a single thread might be viable, too. Maybe compute the sequence for a pair of numbers in parallel, since each one only takes a couple registers, and they can all update the same max / maxi. This creates more instruction-level parallelism.



The trick is deciding whether to wait until all the n values have reached 1 before getting another pair of starting n values, or whether to break out and get a new start point for just one that reached the end condition, without touching the registers for the other sequence. Probably it's best to keep each chain working on useful data, otherwise you'd have to conditionally increment its counter.







You could maybe even do this with SSE packed-compare stuff to conditionally increment the counter for vector elements where n hadn't reached 1 yet. And then to hide the even longer latency of a SIMD conditional-increment implementation, you'd need to keep more vectors of n values up in the air. Maybe only worth with 256b vector (4x uint64_t).



I think the best strategy to make detection of a 1 "sticky" is to mask the vector of all-ones that you add to increment the counter. So after you've seen a 1 in an element, the increment-vector will have a zero, and +=0 is a no-op.



Untested idea for manual vectorization



# starting with YMM0 = [ n_d, n_c, n_b, n_a ]  (64-bit elements)

# ymm4 = _mm256_set1_epi64x(1): increment vector
# ymm5 = all-zeros: count vector

.inner_loop:
vpaddq ymm1, ymm0, xmm0
vpaddq ymm1, ymm1, xmm0
vpaddq ymm1, ymm1, set1_epi64(1) # ymm1= 3*n + 1. Maybe could do this more efficiently?

vprllq ymm3, ymm0, 63 # shift bit 1 to the sign bit


vpsrlq ymm0, ymm0, 1 # n /= 2

# There may be a better way to do this blend, avoiding the bypass delay for an FP blend between integer insns, not sure. Probably worth it
vpblendvpd ymm0, ymm0, ymm1, ymm3 # variable blend controlled by the sign bit of each 64-bit element. I might have the source operands backwards, I always have to look this up.

# ymm0 = updated n in each element.

vpcmpeqq ymm1, ymm0, set1_epi64(1)
vpandn ymm4, ymm1, ymm4 # zero out elements of ymm4 where the compare was true


vpaddq ymm5, ymm5, ymm4 # count++ in elements where n has never been == 1

vptest ymm4, ymm4
jnz .inner_loop
# Fall through when all the n values have reached 1 at some point, and our increment vector is all-zero

vextracti128 ymm0, ymm5, 1
vpmaxq .... crap this doesn't exist
# Actually just delay doing a horizontal max until the very very end. But you need some way to record max and maxi.



You can and should implement this with intrinsics, instead of hand-written asm.






Algorithmic / implementation improvement:



Besides just implementing the same logic with more efficient asm, look for ways to simplify the logic, or avoid redundant work. e.g. memoize to detect common endings to sequences. Or even better, look at 8 trailing bits at once (gnasher's answer)



@EOF points out that tzcnt (or bsf) could be used to do multiple n/=2 iterations in one step. That's probably better than SIMD vectorizing, because no SSE or AVX instruction can do that. It's still compatible with doing multiple scalar ns in parallel in different integer registers, though.




So the loop might look like this:



goto loop_entry;  // C++ structured like the asm, for illustration only
do {
n = n*3 + 1;
loop_entry:
shift = _tzcnt_u64(n);
n >>= shift;
count += shift;

} while(n != 1);


This may do significantly fewer iterations, but variable-count shifts are slow on Intel SnB-family CPUs without BMI2. 3 uops, 2c latency. (They have an input dependency on the FLAGS because count=0 means the flags are unmodified. They handle this as a data dependency, and take multiple uops because a uop can only have 2 inputs (pre-HSW/BDW anyway)). This is the kind that people complaining about x86's crazy-CISC design are referring to. It makes x86 CPUs slower than they would be if the ISA was designed from scratch today, even in a mostly-similar way. (i.e. this is part of the "x86 tax" that costs speed / power.) SHRX/SHLX/SARX (BMI2) are a big win (1 uop / 1c latency).



It also puts tzcnt (3c on Haswell and later) on the critical path, so it significantly lengthens the total latency of the loop-carried dependency chain. It does remove any need for a CMOV, or for preparing a register holding n>>1, though. @Veedrac's answer overcomes all this by deferring the tzcnt/shift for multiple iterations, which is highly effective (see below).



We can safely use BSF or TZCNT interchangeably, because n can never be zero at that point. TZCNT's machine-code decodes as BSF on CPUs that don't support BMI1. (Meaningless prefixes are ignored, so REP BSF runs as BSF).



TZCNT performs much better than BSF on AMD CPUs that support it, so it can be a good idea to use REP BSF, even if you don't care about setting ZF if the input is zero rather than the output. Some compilers do this when you use __builtin_ctzll even with -mno-bmi.




They perform the same on Intel CPUs, so just save the byte if that's all that matters. TZCNT on Intel (pre-Skylake) still has a false-dependency on the supposedly write-only output operand, just like BSF, to support the undocumented behaviour that BSF with input = 0 leaves its destination unmodified. So you need to work around that unless optimizing only for Skylake, so there's nothing to gain from the extra REP byte. (Intel often goes above and beyond what the x86 ISA manual requires, to avoid breaking widely-used code that depends on something it shouldn't, or that is retroactively disallowed. e.g. Windows 9x's assumes no speculative prefetching of TLB entries, which was safe when the code was written, before Intel updated the TLB management rules.)



Anyway, LZCNT/TZCNT on Haswell have the same false dep as POPCNT: see this Q&A. This is why in gcc's asm output for @Veedrac's code, you see it breaking the dep chain with xor-zeroing on the register it's about to use as TZCNT's destination, when it doesn't use dst=src. Since TZCNT/LZCNT/POPCNT never leave their destination undefined or unmodified, this false dependency on the output on Intel CPUs is purely a performance bug / limitation. Presumably it's worth some transistors / power to have them behave like other uops that go to the same execution unit. The only software-visible upside is in the interaction with another microarchitectural limitation: they can micro-fuse a memory operand with an indexed addressing mode on Haswell, but on Skylake where Intel removed the false dependency for LZCNT/TZCNT they "un-laminate" indexed addressing modes while POPCNT can still micro-fuse any addr mode.








@hidefromkgb's answer has a nice observation that you're guaranteed to be able to do one right shift after a 3n+1. You can compute this more even more efficiently than just leaving out the checks between steps. The asm implementation in that answer is broken, though (it depends on OF, which is undefined after SHRD with a count > 1), and slow: ROR rdi,2 is faster than SHRD rdi,rdi,2, and using two CMOV instructions on the critical path is slower than an extra TEST that can run in parallel.




I put tidied / improved C (which guides the compiler to produce better asm), and tested+working faster asm (in comments below the C) up on Godbolt: see the link in @hidefromkgb's answer. (This answer hit the 30k char limit from the large Godbolt URLs, but shortlinks can rot and were too long for goo.gl anyway.)



Also improved the output-printing to convert to a string and make one write() instead of writing one char at a time. This minimizes impact on timing the whole program with perf stat ./collatz (to record performance counters), and I de-obfuscated some of the non-critical asm.






@Veedrac's code



I got a very small speedup from right-shifting as much as we know needs doing, and checking to continue the loop. From 7.5s for limit=1e8 down to 7.275s, on Core2Duo (Merom), with an unroll factor of 16.




code + comments on Godbolt. Don't use this version with clang; it does something silly with the defer-loop. Using a tmp counter k and then adding it to count later changes what clang does, but that slightly hurts gcc.



See discussion in comments: Veedrac's code is excellent on CPUs with BMI1 (i.e. not Celeron/Pentium)


No comments:

Post a Comment

plot explanation - Why did Peaches&#39; mom hang on the tree? - Movies &amp; TV

In the middle of the movie Ice Age: Continental Drift Peaches' mom asked Peaches to go to sleep. Then, she hung on the tree. This parti...