How quickly can you remove spaces from a string?

http://lemire.me/blog/2017/01/20/how-quickly-can-you-remove-spaces-from-a-string/

by deafcalculus • 10 years ago

201 108 10 years ago

lemire.me

et1337• 10 years ago

Cached text-only version: http://webcache.googleusercontent.com/search?q=cache%3Ahttp%...

octo_t• 10 years ago

So using 128bits instructions would imply you had 'words' which were over 16 characters long on average, right?

The average (english) word is ~5 characters long, so most of the time, you'd be forced to check anyway.

countryqt30• 10 years ago

You are so right - everything depends on the dataset. Now after all, just skipping 4 characters (1 byte, 32 bit arch) actually seems VERY reasonable.

Maybe do it log> 1) check if you can jump 16, 2) check if you can jump 8, 3) check if you can jump 4, execute

nkurz5• 10 years ago

This is a good instinct, but doesn't work well in practice unless the prevalence of spaces is very low. And if it's very low, you only need the largest check followed by a single fallback. The problem is that mispredicted branches (for example, 'if statements' that require different code paths without a clear pattern) are relatively expensive.

The scalar operations of reading a character, checking whether it's a space (0x20), and writing it to an output can often be done in a single cycle (the processor is 'superscalar'). A mispredicted branch costs about 15 cycles. Thus for simple tasks like this, if the average distance between spaces is 16 or less, you are likely better off with the simpler straightforward approach.

drdrey• 10 years ago

Not all strings are made of English text

sfrailsdev• 10 years ago

you don't even have to go to non english, just utf-8 has stuff like mathematic spaces and spaces of different em sizes. http://perldoc.perl.org/perlrecharclass.html has a reasonably good list of non ascii spaces, though I'm not sure where to dig through for locale specific lists.

That said, while it may not be a solution for every case, it's a solution for the common case and a starting point for other cases, and thus pretty nifty and potentially useful.

criddell• 10 years ago

Is there a clever way to remove all the space characters?

https://www.cs.tut.fi/~jkorpela/chars/spaces.html

gravypod• 10 years ago

"ZERO WIDTH NO-BREAK SPACE foobar No width (the character is invisible)"

Time to insert this into some of my friends code.

squeaky-clean2• 10 years ago

I've actually joked with colleagues about this, and often it doesn't work. Try using the U+FEFF in Javascript, in Chrome it's taken as a space (but the console will mark it as a red circle, not an invisible no-width character), and Firefox refuses to even allow it to be entered.

You can also use a character and combining character to make a symbol that looks exactly like a semicolon, but technically isn't. Browsers treat it as a semicolon anyways. I wanted to be able to do `var ; = "foobar";`

Maybe it will sneak by other languages that allow unicode in the code? Though I don't know of many.

pjscott2• 10 years ago

If you use the naive algorithm at the top of the page but iterate over UTF-8 code points rather than bytes -- which is straightforward, BTW -- you will get some cleverness automatically in the compiler's implementation of the switch statement. I wrote a switch to handle the Unicode "is it space?" function, with the values from the table you linked, and compiled with "clang -Os". You know what it did?

It generated a friggin' binary search tree! Some of the leaves were a sequence of straight-line comparisons, because that's more compact, but the higher levels were all a bunch of "if (c > 0x167F) { ... }" sort of code. At one point it subtracts 8192 from something and then compares with 12, and I think this is because the x86 instruction encoding is shorter and the compiler knows that the register won't be needed again along either of the code paths from that point.

Compilers are amazing sometimes.

userbinator• 10 years ago

Compiling switches into binary (and even sometimes trinary) trees is an old technique that's been around since at least the days of 16-bit DOS... they will usually choose between a single-level lookup, double-level lookup, or tree depending on sparsity and options used.

Subtraction and addition implicitly sets the flags, so you can generate very small code with inc/dec (1-byte instructions), like this:

    dec ax
    jz ax_was_zero
    dec ax
    jz ax_was_one
    ...

mnarayan01• 10 years ago

I feel like you have to at least include non-breaking spaces if you're going to say you're removing spaces from UTF-8 strings.

dsp12342• 10 years ago

This code will work on all UTF-8 encoded strings… which is the bulk of the strings found on the Internet if you consider that UTF-8 is a superset of ASCII.

Note that the first set of code (and possibly the rest), only work as the space, newline, and carriage return are the 7 bit ASCII set is included in UTF-8. However, the extended 8-bit ASCII set is not, but is often included when people speak of ASCII. So for example, if the request was to remove all "Copyright Sign" symbols, which is U00A9, it would not work correctly. The UTF-8 encoding for this symbol is 0xC2 0xA9, but the code only works on individual bytes, so it would remove the A9 byte, leaving a C2 byte and then whatever byte came next. Additionally, it would hit other UTF-8 characters like the "Greek Capital Letter Omega" (Ω which is encoded in UTF-8 as 0xCE 0xA9)

tl;dr Only works for the 7-bit ASCII set, but not the common extended 8-bit ASCII sets

delinka2• 10 years ago

Do you mean to say that TFA makes the incorrect assumption that UTF-8 is a single byte encoding?

averell• 10 years ago

No, the article states the method will correctly remove the mentioned whitespace characters even from UTF-8 text, because they are part of the 7 bit subset this method works on.

MereInterest• 10 years ago

One neat thing about UTF-8 is that all bytes that compose non-ASCII characters will have the high bit set. ASCII values range from 0-127, so no ASCII values will ever have the high bit set. Therefore, a byte that corresponds to an ASCII value is guaranteed to be the same character as the corresponding ASCII encoding.

ajross• 10 years ago

It's sort of a semantic argument, but I've never once heard anyone refer to "ASCII" to refer to anything other than the 7-bit standard. There is no "extended 8-bit ASCII set". That's a blanket term (though in casual use, "code page" is more typical) for any of the literally dozens of 8-bit character encodings that overlap with ASCII in the bottom half of the encoding space.

Basically the blog post was correct and precise: it removes ASCII characters from UTF-8. There is no elaboration needed.

Again, though, the post was about using SIMD primitives to optimize what looks like a scalar problem.

dragonwriter• 10 years ago

A number of the 8-bit sets that embedded the ASCII 7-bit set were informally referred to as ASCII especially during the 1980s; this may have been encouraged by microcomputer BASIC interpreters using the ASC() function (named for ASCII) to return the character code of a character in the system code set, which was often an 8-bit set that embedded ASCII.

Sharlin4• 10 years ago

You would be surprised. Referring to any of the most common 8-bit encodings as "extended ASCII" or even just "ASCII" has been very common in my experience.

dsp12342• 10 years ago

I mention it because I've seen it happen multiple times over my career.

robterrell• 10 years ago

I'm guessing you're just not old enough to have heard anyone refer to extended ascii?

https://en.wikipedia.org/wiki/Extended_ASCII

If you used Macs in the 80s and 90s and shared files with PC users, you heard about it all the time, because high-ascii characters would map incorrectly across the platforms.

delinka2• 10 years ago

The article does specifically limit itself to ' ', '\n' and '\r' and does not say "all whitespace."

IshKebab2• 10 years ago

I think because of UTF-8 most programmers know that ASCII is 7-bit. What you call '8-bit ASCII' is more commonly called Latin-1.

laumars2• 10 years ago

Latin-1 is region specifc. Windows Code Pages and ISO-8859 (depending on the era) would be better terms to use in your example.

austincheney2• 10 years ago

Isn't this the very kind of thing Regular Expressions were created to solve?

* Remove spaces - myString.replace(/\u0020+/g, "");

* Remove common line terminators - myString.replace(/(\r|\n)+/g, "");

* Remove all white space - myString.replace(/\s+/g, "");

SonicSoul• 10 years ago

the post is not so much about most convenient way to remove spaces, but about efficiency of such an operation.

austincheney2• 10 years ago

Is regex execution considered efficient? I have always been impressed with its speed.

chc• 10 years ago

Regular expressions were created to do this, but your general string replace function with your general regular expression library might not be the fastest way, which is what this post is about.

hcrisp• 10 years ago

Not claiming to be the fastest, but here are two Python solutions suggested by Dave Beazley [0], plus one I extended using compiled regular expressions:

    # string replacement
    s = ' hello  world \n'
    s.replace(' ', '')
    # 'helloworld'

    # regular expression
    import re
    re.sub('\s+', '', s)
    # 'helloworld'

    # compiled regular expression
    pat = re.compile('\s+')
    pat.sub('', s)
    # 'helloworld'

[0] https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&c...

niccl• 10 years ago

my first python thought was

  s = ' hello  world \n'
  s = ''.join(s.split())

again probably not fast, but simple

justinhj• 10 years ago

Scala: val nospaces = "a string with spaces" filter { _ != ' ' }

Darge• 10 years ago

Does anyone know how to do such a benchmark?

sp332• 10 years ago

Check out the linked Github repo.

dzdt2• 10 years ago

Somehow the article fails to mention that the speed will depend heavily on the input string: both its size and distribution of whitespace characters.

Even assuming the question is for the limit of very long strings, the distribution makes a huge difference. Natural English has spaces on average every 5.1 characters [1], so using multi-character tests to speed up the case of runs of 8 or more characters without a space will probably slow it down, not speed it up!

[1]https://arxiv.org/pdf/1208.6109

nkurz5• 10 years ago

using multi-character tests to speed up the case of runs of 8 or more characters without a space will probably slow it down, not speed it up!

Yes, although there is the option of combining the vectorized comparisons with a branchless approach. I modified his code to do this, and got what seemed to be a flat .40 cycles per character independent of input, which is about twice the speed Daniel illustrated on the input with 3% spaces. I'm sure it can be made even faster (what would the limiting factor be?), but I think this shows that a multi-character approach is faster on all input than would be possible with any single-character approach?

rjeli• 10 years ago

Could you share your code?

hyperpape• 10 years ago

Quoting: "It is going to be faster as long as most blocks of eight characters do not contain any white space."

Your point is still helpful because it seems that not having whitespace is probably less common, but there is a mention.

dzdt2• 10 years ago

Fair enough. I should have said "fails to emphasize" not "fails to mention".

wmil• 10 years ago

Perhaps he's using it on authors known to be grandiloquent, loquacious, and sesquipedalian.

dekhn• 10 years ago

you forgot mellifluous

btilly• 10 years ago

For natural language, yes.

But log files tend to have a very different distribution of whitespace.

saretired2• 10 years ago

The “optimized” functions have a bug when the number of remaining bytes is less than the block size.

nkurz5• 10 years ago

Certainly possible, but are you looking at the excerpts in his blog post, or the actual code that he links to? https://github.com/lemire/despacer/blob/master/include/despa...

saretired2• 10 years ago

I'm looking at the excepts in his post, which are broken.

ebbv• 10 years ago

Optimizing by hand is rarely the fastest thing to do nowadays. I wonder how the original naive approach fairs with all optimizations turned on for gcc and with realistic input?

Too2• 10 years ago

The optimization flags used should be good enough.

https://github.com/lemire/despacer/blob/master/Makefile

CFLAGS = -fPIC -std=c99 -O3 -march=native

aib• 10 years ago

I wonder how much of a difference using a separate destination buffer would make. Apart from some memory/cache/invalidation/magic thing that I'm not certain might occur, it would allow one to use the extended instructions used for scanning.

fisherjeff• 10 years ago

Adding lookup tables to the naïve implementation gave me a ~3x speedup with virtually no extra effort. Branch penalties are a real killer.

twoodfin• 10 years ago

You have to be careful with lookup tables. They can be impressive each at a time when microbenchmarking a single loop, but in the aggregate hurt performance by crowding other useful data out of limited cache space.

nogracias• 10 years ago

6.5MB for the tables to drive this faster de-spacer? No thank you. It would blow out the L1 cache and make your program slower.

https://raw.githubusercontent.com/lemire/despacer/master/inc...

jnordwick3• 10 years ago

My pet peeve: using table lookups so a benchmark show faster when in reality your L1 cache is going to be stomped on. Not only will you be waiting on the L1 cache to populate but you also evict all the useful data.

exDM694• 10 years ago

I've seen this elsewhere too. For relatively mundane task (I think it was Morton code conversion), a giant lookup table was constructed and it gave a nice 2x or 4x performance improvement.

But no application will ever be doing only string despacing or Morton codes, so the "fast" lookup table algorithm will make everything else slower by evicting good cache lines. And once something else runs and evicts the lookup tables, the next run will be slow again.

brianpgordon• 10 years ago

This reminded me of spray-json's JsonParser. It has a bit of Scala code to seek past whitespace:

  @tailrec private def ws(): Unit =
    // fast test whether cursorChar is one of " \n\r\t"
    if (((1L << cursorChar) & ((cursorChar - 64) >> 31) & 0x100002600L) != 0L) { advance(); ws() }

https://github.com/spray/spray-json/blob/765c83248e0bbe867dd...

hota_mazi• 10 years ago

The compiler needs to be extremely clever on @tailrec for not blowing up the L1 cache on this code.

To me, this looks like a perfect example of a developer thinking they're very smart while they're actually writing code that's worse that its naive counterpart version (a good summary of Scala in my experience).

jnordwick3• 10 years ago

Can you please explain why it would blow up the L1 cache?

computator• 10 years ago

What strikes me is how many sites won't accept a login name, credit card number, phone number, or other field with leading or trailing spaces. This for something where the speed of the code doesn't matter at all. I can't think of any explanation other than incredible laziness or incompetence for not stripping off spaces.

You might not notice this if you use a password manager or browser autofill, but it's a lot of sites, including companies like major airlines for example.

Never mind that -- it's just a miracle when you can enter a credit card number formatted as 4123 4567 8901 2345 rather than squished together.

edblarney• 10 years ago

You hit the issue in a roundabout way:

The very premise of the article 'how quickly can we remove whitespaces' is rooted in the intellectual foundations of CS. As a culture, we are are obsessive about 'performance'.

This is because back in the day, it's always been important - and even today 'under the hood' it's always important. And of course there are situations in which it's still important (complex algs, limitations of mobile devices).

But in reality - these things are never the issue.

The 'issue' is the pragmatic application of basic algorithms to do a number of basic things elegantly, which together form the foundation of a good user experience.

Yes - the issue of 'no spaces' in card numbers etc. is a clumsy thing, and it's laziness by developers.

Also - things like 'performance' are objectively measurable, you can get cool data for it etc..

A 'bad experience' is sometimes difficult to define.

open-source-ux• 10 years ago

As a culture, we are are obsessive about 'performance'.

My purely anecdotal impression is quite the opposite. Speed of delivery and convenience for the developer (not the end user) seem to be the norm.

Frameworks, scripting languages, browser-based desktop and web apps: none of these have the characteristics of being small, nimble, lightweight, or performant. They certainly make life easier for the developer. Whether users get a 'good experience' out of the end result is open to debate.

tjalfi• 10 years ago

I have an excellent example of this.

I worked at a small technical training company in early 2001. We had an in house application which was used for scheduling classes. If you entered a trailing space the program would beep 3 times and display a dialog box warning you in capital letters that trailing spaces would corrupt the database. The button to close the dialog box was labeled "I understand and will obey". It probably took the developer more time to create the dialog box than to sanitize all inputs.

johnnyb9• 10 years ago

Going to start using this as an interview question!

webo• 10 years ago

Please don't. Especially for a typical software engineering position.

stelund• 10 years ago

Maybe the asm to do scan for byte in memory is an alternative. Repne scab.

It won't find all characters in a single scan. But maybe do 3 passes over a buffer which fits in cache.

cowardlydragon• 10 years ago

ASCII/1 byte characters? I stopped reading then.

UNICODE or GTFO.

Should be embarassingly parallel for a contiguous array.

foxhill2• 10 years ago

this is in no way embarrassingly parallel at all! in fact, it's entirely non trivial to parallelise, writing the output is an inhertintly a serial operation.

mrkgnao• 10 years ago

Could you not do something like chunk the input and operate on the pieces in parallel? I'd think you could use something like a rope (à la text editors) to support the concurrent writes.

(I may or may not have forgotten the differences between concurrency and parallelism.)

ascotan• 10 years ago

I guess you could write CUDA code to do this on a GPU, but then the question becomes why? :/

castratikron• 10 years ago

I don't think this particular problem would be well suited for a GPU. GPUs are not vector processors, they're matrix processors, and the problem that OP is describing is a vector problem. Basically this problem would be limited by the memory bandwidth of a GPU, which isn't a good thing: https://crd.lbl.gov/departments/computer-science/PAR/researc...

antonmks• 10 years ago

Ok, here it is : https://gist.github.com/antonmks/17e0c711d41fb07d1b6f3ada3f5...

0.25 cycles/byte using GTX 1080.

mkagenius• 10 years ago

The Why has multiple answers - one is human curiosity for possibilities of better result (which is the sole cause of all the technology we see around us).

chrismorgan• 10 years ago

A variant of this problem yields further interesting possibilities: if you’re trying to remove control characters like CR and LF, but replacing them with whitespace would be acceptable. That way you can work on it in-place, without needing to copy memory or allocate or anything like that.

Too2• 10 years ago

Someone file a compiler bug? 14x difference between the readable code and the optimized code is a lot. The first code is extremely straightforward, you shouldn't have deal with that SIMD mess manually.

Sharlin4• 10 years ago

Failure to optimize is never a bug, by definition. Automatic vectorization of sequential code is a very difficult problem in general. Especially when your code still depends on testing single bytes in the input and vectorizing that is only possible using some very clever bit twiddling.

exDM694• 10 years ago

Can you come up with a vectorization optimizer pass that could do what the OP did in the article?

This isn't a compiler bug or a shortcoming, it's a non-trivial optimization. Especially given the aliasing between the input and output (ie. it's an in-place algorithm) this is going to be a very difficult optimization.

Annatar• 10 years ago

  awk '
  {
    gsub(/[[:blank:]\015]/, "");
    printf("%s", $0);
  }' input | tee output

stripping out LF isn't necessary because AWK does it on every record automatically. For even more speed, the code can be translated into ANSI C and compiled with awka[1] using an optimizing C compiler.

[1] http://awka.sourceforge.net/index.html

jakobegger• 10 years ago

Fun story: When I implemented the syntax highlighting feature in Sequel Pro, I needed a way to determine the string length of UTF-8 encoded strings. I didn't find a function available on macOS that worked directly on a char*, so I googled and found a really simple UTF-8 strlen function. It was easy to understand and very fast. I think it was this: http://canonical.org/~kragen/strlen-utf8.html

I committed my syntax highlighting code, and a few days later someone had replaced the simple UTF8 strlen function with the really long vectorised version from this page: http://www.daemonology.net/blog/2008-06-05-faster-utf8-strle...

But the funny thing is, that the supposedly fast vectorised strlen was optimised for very long strings. The benchmarks shows results for megabytes of texts. But we measured the length of tokens, usually only a few characters long, so in most cases the new function was actually slower!

I was a very junior dev, didn't want to piss off anybody, and the strlen wasn't in the hot path anyway, so I didn't say anything. But I was a bit sad, that my easy-to-read code was replaced by such a monstrosity.

What's my point? Before you go and use these functions in your code, profile your code to see if it would actually affect performance.

exDM694• 10 years ago

I am surprised to see any speedup in this! I'd expect something trivial like this to be completely memory bound with the CPU sitting almost idle waiting for bytes coming in from memory.

Looking at the benchmark code, this is using rdtsc to read the CPU time stamp counter. That does not take waiting for memory into account, does it?

I wonder if there's a difference when measured in wall clock time. It's still somewhat beneficial to have the CPU work efficiently to give an opportunity for hyperthreading to take place when waiting for memory.

If you really wanted to make something like this faster, you should focus on cache utilization and make use of prefetching instructions. x86 has pretty bad prefetching instructions and pretty good speculative fetching, so don't expect massive speedups but on ARM or Aarch64, you have a finer grained control over cache prefetching (L1 and L2 separately) and you could see much bigger differences.

As for benchmarking this kind of problems: you obviously want to measure real world performance, so you need wall clock time as well as time stamp counter, but I'd look for optimization clues in "perf stat" and other CPU perf counters, with an emphasis on cache misses and branch mispredictions.

The figure you should be staring at is the total throughput of the algorithm, measured in gigabytes per second. You should be getting figures close to the memory bandwidth available (25-50 GB/s depending on CPU and memory).

edit: I measured the wall clock time with clock_gettime before/after all the repeats (using a megabyte sized buffer) and there is indeed no significant difference, here's my results:

    memcpy(tmpbuffer,buffer,N):  0.122945 cycles / ops 1495907352 nsec (1.495907 sec) 
    countspaces(buffer, N):  3.657322 cycles / ops 1544915395 nsec (1.544915 sec) 
    despace(buffer, N):  6.521193 cycles / ops 1621204460 nsec (1.621204 sec) 
    faster_despace(buffer, N):  1.721657 cycles / ops 1500507217 nsec (1.500507 sec) 
    despace64(buffer, N):  3.595031 cycles / ops 1544993649 nsec (1.544994 sec) 
    despace_to(buffer, N, tmpbuffer):  6.307885 cycles / ops 1615101563 nsec (1.615102 sec) 
    avx2_countspaces(buffer, N):  0.190992 cycles / ops 1460961459 nsec (1.460961 sec) 
    avx2_despace(buffer, N):  5.750583 cycles / ops 1615971010 nsec (1.615971 sec) 
    sse4_despace(buffer, N):  0.985002 cycles / ops 1482901389 nsec (1.482901 sec) 
    sse4_despace_branchless(buffer, N):  0.338737 cycles / ops 1460874704 nsec (1.460875 sec) 
    sse4_despace_trail(buffer, N):  1.950657 cycles / ops 1502268447 nsec (1.502268 sec) 
    sse42_despace_branchless(buffer, N):  0.562246 cycles / ops 1468638389 nsec (1.468638 sec) 
    sse42_despace_branchless_lookup(buffer, N):  0.624913 cycles / ops 1472445127 nsec (1.472445 sec) 
    sse42_despace_to(buffer, N,tmpbuffer):  1.747046 cycles / ops 1507705780 nsec (1.507706 sec)

Here's the diff to the original: http://pasteall.org/208511

edit2: surprisingly, Clang is about 10% slower than GCC in my experiments.

nkurz5• 10 years ago

Thanks for looking into this! I know that Daniel appreciates feedback, but rarely reads HN, so an email to him or blog comment with your results might be helpful.

Looking at the benchmark code, this is using rdtsc to read the CPU time stamp counter. That does not take waiting for memory into account, does it?

On modern Intel processors, the "time stamp counter" is monotonically increasing at a constant rate, so it does take memory latency into account. On many Linux systems (including the one used here) clock_gettime() uses the same underlying clock source, so there should be no difference in accuracy for long measurements. The CPUID-RDTSC/RDSTCP-CPUID pattern used here has the advantage of a somewhat lower and significantly more predictable overhead, which helps when measuring shorter events.

I'd expect something trivial like this to be completely memory bound with the CPU sitting almost idle waiting for bytes coming in from memory.

If I remember the numbers right, the Skylake processor this is running on can read about 64B per cycle if the source is L1, about 24B per cycle if the source is L3, and about 6B per cycle from main memory. Using the existing RDTSC framework, I get equal speeds at L3 size, and still get sub .4B/cycle coming from main memory.

I measured the wall clock time with clock_gettime before/after all the repeats (using a megabyte sized buffer) and there is indeed no significant difference

I agree that testing on larger buffers would be informative (as would testing different ratios of whitespace) but I don't think your approach is capturing what you think it is. The macro being used for time measurement uses a different random input for each iteration (look at 'pre'), and the unoptimized time of initializing this dominates the clock time. So while I think your test is worthwhile, I think you need a better way to perform the measurement. I think you'll see that RDTSC maps exactly to wall time in this case, but surprises are definitely possible.

I'd look for optimization clues in "perf stat" and other CPU perf counters, with an emphasis on cache misses and branch mispredictions.

Yes, although as with wall time, one needs to be sure to measure only on the section of code that one is optimizing. Perf makes this difficult, so something like "likwid" with an API that allows profiling fragments of code would be required. I haven't done that yet for this, but at the faster speeds (sse4_despace_branchless) I don't think there are any cache misses or branch mispredictions in the code of interest.

surprisingly, Clang is about 10% slower than GCC in my experiments

My surprise was the opposite. Clang was 25% faster than GCC and ICC on the L1- and L3-sized vectorized branchless (.25 cycles/byte versus .33 cycles/byte, although the code I'm running is not quite what's on Github). Most of this benefit seems to be because clang unrolls 2x to reduce loop overhead.

exDM694• 10 years ago

I dropped a comment on his blog and a link to this discussion...

And how sloppy of me! I didn't notice the `pre;` in the code consuming most of the time (to be fair it's just 4 characters and I didn't put too much time to it). When I move that outside of the loop, before the timer, I get results that show improvement with the optimized version.

And indeed it looks like rdtsc is giving similar figures to clock_gettime now. I falsely presumed that it counts retired instructions.

I'm still surprised to see a speedup, and how badly the original version is performing.

My guess is that the conditional stores are poison to the CPU pipelines. The 64 bit version gets most of the performance out of it already, I presume it's because of the more efficient memory usage pattern.

I can't edit my earlier post any more, I'd correct it if I could.

Clang:

    memcpy(tmpbuffer,buffer,N):  0.000000 cycles / ops 65232 nsec (0.000065 sec) 
    countspaces(buffer, N):  0.000000 cycles / ops 65176 nsec (0.000065 sec) 
    despace(buffer, N):  4.547286 cycles / ops 7654310198 nsec (7.654310 sec) 
    faster_despace(buffer, N):  1.582721 cycles / ops 2651677500 nsec (2.651678 sec) 
    despace64(buffer, N):  0.583952 cycles / ops 1025215835 nsec (1.025216 sec) 
    despace_to(buffer, N, tmpbuffer):  0.000000 cycles / ops 63847 nsec (0.000064 sec) 
    avx2_countspaces(buffer, N):  0.000000 cycles / ops 63697 nsec (0.000064 sec) 
    avx2_despace(buffer, N):  0.307253 cycles / ops 602528061 nsec (0.602528 sec) 
    sse4_despace(buffer, N):  0.310504 cycles / ops 534542967 nsec (0.534543 sec) 
    sse4_despace_branchless(buffer, N):  0.353851 cycles / ops 594652080 nsec (0.594652 sec) 
    sse4_despace_trail(buffer, N):  0.314221 cycles / ops 562439990 nsec (0.562440 sec) 
    sse42_despace_branchless(buffer, N):  0.608811 cycles / ops 1020576734 nsec (1.020577 sec) 
    sse42_despace_branchless_lookup(buffer, N):  0.608494 cycles / ops 1020062750 nsec (1.020063 sec) 
    sse42_despace_to(buffer, N,tmpbuffer):  1.779908 cycles / ops 2983955982 nsec (2.983956 sec)

GCC:

    memcpy(tmpbuffer,buffer,N):  0.285702 cycles / ops 489599835 nsec (0.489600 sec) 
    countspaces(buffer, N):  0.000000 cycles / ops 63995 nsec (0.000064 sec) 
    despace(buffer, N):  4.751018 cycles / ops 8014809122 nsec (8.014809 sec) 
    faster_despace(buffer, N):  1.718575 cycles / ops 2898082416 nsec (2.898082 sec) 
    despace64(buffer, N):  0.883421 cycles / ops 1560479651 nsec (1.560480 sec) 
    despace_to(buffer, N, tmpbuffer):  6.424313 cycles / ops 10892716258 nsec (10.892716 sec) 
    avx2_countspaces(buffer, N):  0.031227 cycles / ops 52369789 nsec (0.052370 sec) 
    avx2_despace(buffer, N):  0.315633 cycles / ops 627793484 nsec (0.627793 sec) 
    sse4_despace(buffer, N):  0.319739 cycles / ops 554173689 nsec (0.554174 sec) 
    sse4_despace_branchless(buffer, N):  0.366240 cycles / ops 615638316 nsec (0.615638 sec) 
    sse4_despace_trail(buffer, N):  0.318973 cycles / ops 572262343 nsec (0.572262 sec) 
    sse42_despace_branchless(buffer, N):  0.638114 cycles / ops 1070772787 nsec (1.070773 sec) 
    sse42_despace_branchless_lookup(buffer, N):  0.637988 cycles / ops 1069642163 nsec (1.069642 sec) 
    sse42_despace_to(buffer, N,tmpbuffer):  1.768081 cycles / ops 2963488990 nsec (2.963489 sec)

How quickly can you remove spaces from a string?

Related Stories