Replacing a 32-bit loop counter with 64-bit introd

2019-01-01 09:48发布

问题:

I was looking for the fastest way to popcount large arrays of data. I encountered a very weird effect: Changing the loop variable from unsigned to uint64_t made the performance drop by 50% on my PC.

The Benchmark

#include <iostream>
#include <chrono>
#include <x86intrin.h>

int main(int argc, char* argv[]) {

    using namespace std;
    if (argc != 2) {
       cerr << \"usage: array_size in MB\" << endl;
       return -1;
    }

    uint64_t size = atol(argv[1])<<20;
    uint64_t* buffer = new uint64_t[size/8];
    char* charbuffer = reinterpret_cast<char*>(buffer);
    for (unsigned i=0; i<size; ++i)
        charbuffer[i] = rand()%256;

    uint64_t count,duration;
    chrono::time_point<chrono::system_clock> startP,endP;
    {
        startP = chrono::system_clock::now();
        count = 0;
        for( unsigned k = 0; k < 10000; k++){
            // Tight unrolled loop with unsigned
            for (unsigned i=0; i<size/8; i+=4) {
                count += _mm_popcnt_u64(buffer[i]);
                count += _mm_popcnt_u64(buffer[i+1]);
                count += _mm_popcnt_u64(buffer[i+2]);
                count += _mm_popcnt_u64(buffer[i+3]);
            }
        }
        endP = chrono::system_clock::now();
        duration = chrono::duration_cast<std::chrono::nanoseconds>(endP-startP).count();
        cout << \"unsigned\\t\" << count << \'\\t\' << (duration/1.0E9) << \" sec \\t\"
             << (10000.0*size)/(duration) << \" GB/s\" << endl;
    }
    {
        startP = chrono::system_clock::now();
        count=0;
        for( unsigned k = 0; k < 10000; k++){
            // Tight unrolled loop with uint64_t
            for (uint64_t i=0;i<size/8;i+=4) {
                count += _mm_popcnt_u64(buffer[i]);
                count += _mm_popcnt_u64(buffer[i+1]);
                count += _mm_popcnt_u64(buffer[i+2]);
                count += _mm_popcnt_u64(buffer[i+3]);
            }
        }
        endP = chrono::system_clock::now();
        duration = chrono::duration_cast<std::chrono::nanoseconds>(endP-startP).count();
        cout << \"uint64_t\\t\"  << count << \'\\t\' << (duration/1.0E9) << \" sec \\t\"
             << (10000.0*size)/(duration) << \" GB/s\" << endl;
    }

    free(charbuffer);
}

As you see, we create a buffer of random data, with the size being x megabytes where x is read from the command line. Afterwards, we iterate over the buffer and use an unrolled version of the x86 popcount intrinsic to perform the popcount. To get a more precise result, we do the popcount 10,000 times. We measure the times for the popcount. In the upper case, the inner loop variable is unsigned, in the lower case, the inner loop variable is uint64_t. I thought that this should make no difference, but the opposite is the case.

The (absolutely crazy) results

I compile it like this (g++ version: Ubuntu 4.8.2-19ubuntu1):

g++ -O3 -march=native -std=c++11 test.cpp -o test

Here are the results on my Haswell Core i7-4770K CPU @ 3.50 GHz, running test 1 (so 1 MB random data):

  • unsigned 41959360000 0.401554 sec 26.113 GB/s
  • uint64_t 41959360000 0.759822 sec 13.8003 GB/s

As you see, the throughput of the uint64_t version is only half the one of the unsigned version! The problem seems to be that different assembly gets generated, but why? First, I thought of a compiler bug, so I tried clang++ (Ubuntu Clang version 3.4-1ubuntu3):

clang++ -O3 -march=native -std=c++11 teest.cpp -o test

Result: test 1

  • unsigned 41959360000 0.398293 sec 26.3267 GB/s
  • uint64_t 41959360000 0.680954 sec 15.3986 GB/s

So, it is almost the same result and is still strange. But now it gets super strange. I replace the buffer size that was read from input with a constant 1, so I change:

uint64_t size = atol(argv[1]) << 20;

to

uint64_t size = 1 << 20;

Thus, the compiler now knows the buffer size at compile time. Maybe it can add some optimizations! Here are the numbers for g++:

  • unsigned 41959360000 0.509156 sec 20.5944 GB/s
  • uint64_t 41959360000 0.508673 sec 20.6139 GB/s

Now, both versions are equally fast. However, the unsigned got even slower! It dropped from 26 to 20 GB/s, thus replacing a non-constant by a constant value lead to a deoptimization. Seriously, I have no clue what is going on here! But now to clang++ with the new version:

  • unsigned 41959360000 0.677009 sec 15.4884 GB/s
  • uint64_t 41959360000 0.676909 sec 15.4906 GB/s

Wait, what? Now, both versions dropped to the slow number of 15 GB/s. Thus, replacing a non-constant by a constant value even lead to slow code in both cases for Clang!

I asked a colleague with an Ivy Bridge CPU to compile my benchmark. He got similar results, so it does not seem to be Haswell. Because two compilers produce strange results here, it also does not seem to be a compiler bug. We do not have an AMD CPU here, so we could only test with Intel.

More madness, please!

Take the first example (the one with atol(argv[1])) and put a static before the variable, i.e.:

static uint64_t size=atol(argv[1])<<20;

Here are my results in g++:

  • unsigned 41959360000 0.396728 sec 26.4306 GB/s
  • uint64_t 41959360000 0.509484 sec 20.5811 GB/s

Yay, yet another alternative. We still have the fast 26 GB/s with u32, but we managed to get u64 at least from the 13 GB/s to the 20 GB/s version! On my collegue\'s PC, the u64 version became even faster than the u32 version, yielding the fastest result of all. Sadly, this only works for g++, clang++ does not seem to care about static.

My question

Can you explain these results? Especially:

  • How can there be such a difference between u32 and u64?
  • How can replacing a non-constant by a constant buffer size trigger less optimal code?
  • How can the insertion of the static keyword make the u64 loop faster? Even faster than the original code on my collegue\'s computer!

I know that optimization is a tricky territory, however, I never thought that such small changes can lead to a 100% difference in execution time and that small factors like a constant buffer size can again mix results totally. Of course, I always want to have the version that is able to popcount 26 GB/s. The only reliable way I can think of is copy paste the assembly for this case and use inline assembly. This is the only way I can get rid of compilers that seem to go mad on small changes. What do you think? Is there another way to reliably get the code with most performance?

The Disassembly

Here is the disassembly for the various results:

26 GB/s version from g++ / u32 / non-const bufsize:

0x400af8:
lea 0x1(%rdx),%eax
popcnt (%rbx,%rax,8),%r9
lea 0x2(%rdx),%edi
popcnt (%rbx,%rcx,8),%rax
lea 0x3(%rdx),%esi
add %r9,%rax
popcnt (%rbx,%rdi,8),%rcx
add $0x4,%edx
add %rcx,%rax
popcnt (%rbx,%rsi,8),%rcx
add %rcx,%rax
mov %edx,%ecx
add %rax,%r14
cmp %rbp,%rcx
jb 0x400af8

13 GB/s version from g++ / u64 / non-const bufsize:

0x400c00:
popcnt 0x8(%rbx,%rdx,8),%rcx
popcnt (%rbx,%rdx,8),%rax
add %rcx,%rax
popcnt 0x10(%rbx,%rdx,8),%rcx
add %rcx,%rax
popcnt 0x18(%rbx,%rdx,8),%rcx
add $0x4,%rdx
add %rcx,%rax
add %rax,%r12
cmp %rbp,%rdx
jb 0x400c00

15 GB/s version from clang++ / u64 / non-const bufsize:

0x400e50:
popcnt (%r15,%rcx,8),%rdx
add %rbx,%rdx
popcnt 0x8(%r15,%rcx,8),%rsi
add %rdx,%rsi
popcnt 0x10(%r15,%rcx,8),%rdx
add %rsi,%rdx
popcnt 0x18(%r15,%rcx,8),%rbx
add %rdx,%rbx
add $0x4,%rcx
cmp %rbp,%rcx
jb 0x400e50

20 GB/s version from g++ / u32&u64 / const bufsize:

0x400a68:
popcnt (%rbx,%rdx,1),%rax
popcnt 0x8(%rbx,%rdx,1),%rcx
add %rax,%rcx
popcnt 0x10(%rbx,%rdx,1),%rax
add %rax,%rcx
popcnt 0x18(%rbx,%rdx,1),%rsi
add $0x20,%rdx
add %rsi,%rcx
add %rcx,%rbp
cmp $0x100000,%rdx
jne 0x400a68

15 GB/s version from clang++ / u32&u64 / const bufsize:

0x400dd0:
popcnt (%r14,%rcx,8),%rdx
add %rbx,%rdx
popcnt 0x8(%r14,%rcx,8),%rsi
add %rdx,%rsi
popcnt 0x10(%r14,%rcx,8),%rdx
add %rsi,%rdx
popcnt 0x18(%r14,%rcx,8),%rbx
add %rdx,%rbx
add $0x4,%rcx
cmp $0x20000,%rcx
jb 0x400dd0

Interestingly, the fastest (26 GB/s) version is also the longest! It seems to be the only solution that uses lea. Some versions use jb to jump, others use jne. But apart from that, all versions seem to be comparable. I don\'t see where a 100% performance gap could originate from, but I am not too adept at deciphering assembly. The slowest (13 GB/s) version looks even very short and good. Can anyone explain this?

Lessons learned

No matter what the answer to this question will be; I have learned that in really hot loops every detail can matter, even details that do not seem to have any association to the hot code. I have never thought about what type to use for a loop variable, but as you see such a minor change can make a 100% difference! Even the storage type of a buffer can make a huge difference, as we saw with the insertion of the static keyword in front of the size variable! In the future, I will always test various alternatives on various compilers when writing really tight and hot loops that are crucial for system performance.

The interesting thing is also that the performance difference is still so high although I have already unrolled the loop four times. So even if you unroll, you can still get hit by major performance deviations. Quite interesting.

回答1:

Culprit: False Data Dependency (and the compiler isn\'t even aware of it)

On Sandy/Ivy Bridge and Haswell processors, the instruction:

popcnt  src, dest

appears to have a false dependency on the destination register dest. Even though the instruction only writes to it, the instruction will wait until dest is ready before executing.

This dependency doesn\'t just hold up the 4 popcnts from a single loop iteration. It can carry across loop iterations making it impossible for the processor to parallelize different loop iterations.

The unsigned vs. uint64_t and other tweaks don\'t directly affect the problem. But they influence the register allocator which assigns the registers to the variables.

In your case, the speeds are a direct result of what is stuck to the (false) dependency chain depending on what the register allocator decided to do.

  • 13 GB/s has a chain: popcnt-add-popcnt-popcnt → next iteration
  • 15 GB/s has a chain: popcnt-add-popcnt-add → next iteration
  • 20 GB/s has a chain: popcnt-popcnt → next iteration
  • 26 GB/s has a chain: popcnt-popcnt → next iteration

The difference between 20 GB/s and 26 GB/s seems to be a minor artifact of the indirect addressing. Either way, the processor starts to hit other bottlenecks once you reach this speed.


To test this, I used inline assembly to bypass the compiler and get exactly the assembly I want. I also split up the count variable to break all other dependencies that might mess with the benchmarks.

Here are the results:

Sandy Bridge Xeon @ 3.5 GHz: (full test code can be found at the bottom)

  • GCC 4.6.3: g++ popcnt.cpp -std=c++0x -O3 -save-temps -march=native
  • Ubuntu 12

Different Registers: 18.6195 GB/s

.L4:
    movq    (%rbx,%rax,8), %r8
    movq    8(%rbx,%rax,8), %r9
    movq    16(%rbx,%rax,8), %r10
    movq    24(%rbx,%rax,8), %r11
    addq    $4, %rax

    popcnt %r8, %r8
    add    %r8, %rdx
    popcnt %r9, %r9
    add    %r9, %rcx
    popcnt %r10, %r10
    add    %r10, %rdi
    popcnt %r11, %r11
    add    %r11, %rsi

    cmpq    $131072, %rax
    jne .L4

Same Register: 8.49272 GB/s

.L9:
    movq    (%rbx,%rdx,8), %r9
    movq    8(%rbx,%rdx,8), %r10
    movq    16(%rbx,%rdx,8), %r11
    movq    24(%rbx,%rdx,8), %rbp
    addq    $4, %rdx

    # This time reuse \"rax\" for all the popcnts.
    popcnt %r9, %rax
    add    %rax, %rcx
    popcnt %r10, %rax
    add    %rax, %rsi
    popcnt %r11, %rax
    add    %rax, %r8
    popcnt %rbp, %rax
    add    %rax, %rdi

    cmpq    $131072, %rdx
    jne .L9

Same Register with broken chain: 17.8869 GB/s

.L14:
    movq    (%rbx,%rdx,8), %r9
    movq    8(%rbx,%rdx,8), %r10
    movq    16(%rbx,%rdx,8), %r11
    movq    24(%rbx,%rdx,8), %rbp
    addq    $4, %rdx

    # Reuse \"rax\" for all the popcnts.
    xor    %rax, %rax    # Break the cross-iteration dependency by zeroing \"rax\".
    popcnt %r9, %rax
    add    %rax, %rcx
    popcnt %r10, %rax
    add    %rax, %rsi
    popcnt %r11, %rax
    add    %rax, %r8
    popcnt %rbp, %rax
    add    %rax, %rdi

    cmpq    $131072, %rdx
    jne .L14

So what went wrong with the compiler?

It seems that neither GCC nor Visual Studio are aware that popcnt has such a false dependency. Nevertheless, these false dependencies aren\'t uncommon. It\'s just a matter of whether the compiler is aware of it.

popcnt isn\'t exactly the most used instruction. So it\'s not really a surprise that a major compiler could miss something like this. There also appears to be no documentation anywhere that mentions this problem. If Intel doesn\'t disclose it, then nobody outside will know until someone runs into it by chance.

(Update: As of version 4.9.2, GCC is aware of this false-dependency and generates code to compensate it when optimizations are enabled. Major compilers from other vendors, including Clang, MSVC, and even Intel\'s own ICC are not yet aware of this microarchitectural erratum and will not emit code that compensates for it.)

Why does the CPU have such a false dependency?

We can only speculate, but it\'s likely that Intel has the same handling for a lot of two-operand instructions. Common instructions like add, sub take two operands both of which are inputs. So Intel probably shoved popcnt into the same category to keep the processor design simple.

AMD processors do not appear to have this false dependency.


The full test code is below for reference:

#include <iostream>
#include <chrono>
#include <x86intrin.h>

int main(int argc, char* argv[]) {

   using namespace std;
   uint64_t size=1<<20;

   uint64_t* buffer = new uint64_t[size/8];
   char* charbuffer=reinterpret_cast<char*>(buffer);
   for (unsigned i=0;i<size;++i) charbuffer[i]=rand()%256;

   uint64_t count,duration;
   chrono::time_point<chrono::system_clock> startP,endP;
   {
      uint64_t c0 = 0;
      uint64_t c1 = 0;
      uint64_t c2 = 0;
      uint64_t c3 = 0;
      startP = chrono::system_clock::now();
      for( unsigned k = 0; k < 10000; k++){
         for (uint64_t i=0;i<size/8;i+=4) {
            uint64_t r0 = buffer[i + 0];
            uint64_t r1 = buffer[i + 1];
            uint64_t r2 = buffer[i + 2];
            uint64_t r3 = buffer[i + 3];
            __asm__(
                \"popcnt %4, %4  \\n\\t\"
                \"add %4, %0     \\n\\t\"
                \"popcnt %5, %5  \\n\\t\"
                \"add %5, %1     \\n\\t\"
                \"popcnt %6, %6  \\n\\t\"
                \"add %6, %2     \\n\\t\"
                \"popcnt %7, %7  \\n\\t\"
                \"add %7, %3     \\n\\t\"
                : \"+r\" (c0), \"+r\" (c1), \"+r\" (c2), \"+r\" (c3)
                : \"r\"  (r0), \"r\"  (r1), \"r\"  (r2), \"r\"  (r3)
            );
         }
      }
      count = c0 + c1 + c2 + c3;
      endP = chrono::system_clock::now();
      duration=chrono::duration_cast<std::chrono::nanoseconds>(endP-startP).count();
      cout << \"No Chain\\t\" << count << \'\\t\' << (duration/1.0E9) << \" sec \\t\"
            << (10000.0*size)/(duration) << \" GB/s\" << endl;
   }
   {
      uint64_t c0 = 0;
      uint64_t c1 = 0;
      uint64_t c2 = 0;
      uint64_t c3 = 0;
      startP = chrono::system_clock::now();
      for( unsigned k = 0; k < 10000; k++){
         for (uint64_t i=0;i<size/8;i+=4) {
            uint64_t r0 = buffer[i + 0];
            uint64_t r1 = buffer[i + 1];
            uint64_t r2 = buffer[i + 2];
            uint64_t r3 = buffer[i + 3];
            __asm__(
                \"popcnt %4, %%rax   \\n\\t\"
                \"add %%rax, %0      \\n\\t\"
                \"popcnt %5, %%rax   \\n\\t\"
                \"add %%rax, %1      \\n\\t\"
                \"popcnt %6, %%rax   \\n\\t\"
                \"add %%rax, %2      \\n\\t\"
                \"popcnt %7, %%rax   \\n\\t\"
                \"add %%rax, %3      \\n\\t\"
                : \"+r\" (c0), \"+r\" (c1), \"+r\" (c2), \"+r\" (c3)
                : \"r\"  (r0), \"r\"  (r1), \"r\"  (r2), \"r\"  (r3)
                : \"rax\"
            );
         }
      }
      count = c0 + c1 + c2 + c3;
      endP = chrono::system_clock::now();
      duration=chrono::duration_cast<std::chrono::nanoseconds>(endP-startP).count();
      cout << \"Chain 4   \\t\"  << count << \'\\t\' << (duration/1.0E9) << \" sec \\t\"
            << (10000.0*size)/(duration) << \" GB/s\" << endl;
   }
   {
      uint64_t c0 = 0;
      uint64_t c1 = 0;
      uint64_t c2 = 0;
      uint64_t c3 = 0;
      startP = chrono::system_clock::now();
      for( unsigned k = 0; k < 10000; k++){
         for (uint64_t i=0;i<size/8;i+=4) {
            uint64_t r0 = buffer[i + 0];
            uint64_t r1 = buffer[i + 1];
            uint64_t r2 = buffer[i + 2];
            uint64_t r3 = buffer[i + 3];
            __asm__(
                \"xor %%rax, %%rax   \\n\\t\"   // <--- Break the chain.
                \"popcnt %4, %%rax   \\n\\t\"
                \"add %%rax, %0      \\n\\t\"
                \"popcnt %5, %%rax   \\n\\t\"
                \"add %%rax, %1      \\n\\t\"
                \"popcnt %6, %%rax   \\n\\t\"
                \"add %%rax, %2      \\n\\t\"
                \"popcnt %7, %%rax   \\n\\t\"
                \"add %%rax, %3      \\n\\t\"
                : \"+r\" (c0), \"+r\" (c1), \"+r\" (c2), \"+r\" (c3)
                : \"r\"  (r0), \"r\"  (r1), \"r\"  (r2), \"r\"  (r3)
                : \"rax\"
            );
         }
      }
      count = c0 + c1 + c2 + c3;
      endP = chrono::system_clock::now();
      duration=chrono::duration_cast<std::chrono::nanoseconds>(endP-startP).count();
      cout << \"Broken Chain\\t\"  << count << \'\\t\' << (duration/1.0E9) << \" sec \\t\"
            << (10000.0*size)/(duration) << \" GB/s\" << endl;
   }

   free(charbuffer);
}

An equally interesting benchmark can be found here: http://pastebin.com/kbzgL8si
This benchmark varies the number of popcnts that are in the (false) dependency chain.

False Chain 0:  41959360000 0.57748 sec     18.1578 GB/s
False Chain 1:  41959360000 0.585398 sec    17.9122 GB/s
False Chain 2:  41959360000 0.645483 sec    16.2448 GB/s
False Chain 3:  41959360000 0.929718 sec    11.2784 GB/s
False Chain 4:  41959360000 1.23572 sec     8.48557 GB/s


回答2:

I coded up an equivalent C program to experiment, and I can confirm this strange behaviour. What\'s more, gcc believes the 64-bit integer (which should probably be a size_t anyway...) to be better, as using uint_fast32_t causes gcc to use a 64-bit uint.

I did a bit of mucking around with the assembly:
Simply take the 32-bit version, replace all 32-bit instructions/registers with the 64-bit version in the inner popcount-loop of the program. Observation: the code is just as fast as the 32-bit version!

This is obviously a hack, as the size of the variable isn\'t really 64 bit, as other parts of the program still use the 32-bit version, but as long as the inner popcount-loop dominates performance, this is a good start.

I then copied the inner loop code from the 32-bit version of the program, hacked it up to be 64 bit, fiddled with the registers to make it a replacement for the inner loop of the 64-bit version. This code also runs as fast as the 32-bit version.

My conclusion is that this is bad instruction scheduling by the compiler, not actual speed/latency advantage of 32-bit instructions.

(Caveat: I hacked up assembly, could have broken something without noticing. I don\'t think so.)



回答3:

This is not an answer, but it\'s hard to read if I put results in comment.

I get these results with a Mac Pro (Westmere 6-Cores Xeon 3.33 GHz). I compiled it with clang -O3 -msse4 -lstdc++ a.cpp -o a (-O2 get same result).

clang with uint64_t size=atol(argv[1])<<20;

unsigned    41950110000 0.811198 sec    12.9263 GB/s
uint64_t    41950110000 0.622884 sec    16.8342 GB/s

clang with uint64_t size=1<<20;

unsigned    41950110000 0.623406 sec    16.8201 GB/s
uint64_t    41950110000 0.623685 sec    16.8126 GB/s

I also tried to:

  1. Reverse the test order, the result is the same so it rules out the cache factor.
  2. Have the for statement in reverse: for (uint64_t i=size/8;i>0;i-=4). This gives the same result and proves the compile is smart enough to not divide size by 8 every iteration (as expected).

Here is my wild guess:

The speed factor comes in three parts:

  • code cache: uint64_t version has larger code size, but this does not have an effect on my Xeon CPU. This makes the 64-bit version slower.

  • Instructions used. Note not only the loop count, but the buffer is accessed with a 32-bit and 64-bit index on the two versions. Accessing a pointer with a 64-bit offset requests a dedicated 64-bit register and addressing, while you can use immediate for a 32-bit offset. This may make the 32-bit version faster.

  • Instructions are only emitted on the 64-bit compile (that is, prefetch). This makes 64-bit faster.

The three factors together match with the observed seemingly conflicting results.



回答4:

I can\'t give an authoritative answer, but provide an overview of a likely cause. This reference shows pretty clearly that for the instructions in the body of your loop there is a 3:1 ratio between latency and throughput. It also shows the effects of multiple dispatch. Since there are (give-or-take) three integer units in modern x86 processors, it\'s generally possible to dispatch three instructions per cycle.

So between peak pipeline and multiple dispatch performance and failure of these mechanisms, we have a factor of six in performance. It\'s pretty well known that the complexity of the x86 instruction set makes it quite easy for quirky breakage to occur. The document above has a great example:

The Pentium 4 performance for 64-bit right shifts is really poor. 64-bit left shift as well as all 32-bit shifts have acceptable performance. It appears that the data path from the upper 32 bits to the lower 32 bit of the ALU is not well designed.

I personally ran into a strange case where a hot loop ran considerably slower on a specific core of a four-core chip (AMD if I recall). We actually got better performance on a map-reduce calculation by turning that core off.

Here my guess is contention for integer units: that the popcnt, loop counter, and address calculations can all just barely run at full speed with the 32-bit wide counter, but the 64-bit counter causes contention and pipeline stalls. Since there are only about 12 cycles total, potentially 4 cycles with multiple dispatch, per loop body execution, a single stall could reasonably affect run time by a factor of 2.

The change induced by using a static variable, which I\'m guessing just causes a minor reordering of instructions, is another clue that the 32-bit code is at some tipping point for contention.

I know this is not a rigorous analysis, but it is a plausible explanation.



回答5:

I tried this with Visual Studio 2013 Express, using a pointer instead of an index, which sped up the process a bit. I suspect this is because the addressing is offset + register, instead of offset + register + (register<<3). C++ code.

   uint64_t* bfrend = buffer+(size/8);
   uint64_t* bfrptr;

// ...

   {
      startP = chrono::system_clock::now();
      count = 0;
      for (unsigned k = 0; k < 10000; k++){
         // Tight unrolled loop with uint64_t
         for (bfrptr = buffer; bfrptr < bfrend;){
            count += __popcnt64(*bfrptr++);
            count += __popcnt64(*bfrptr++);
            count += __popcnt64(*bfrptr++);
            count += __popcnt64(*bfrptr++);
         }
      }
      endP = chrono::system_clock::now();
      duration = chrono::duration_cast<std::chrono::nanoseconds>(endP-startP).count();
      cout << \"uint64_t\\t\"  << count << \'\\t\' << (duration/1.0E9) << \" sec \\t\"
           << (10000.0*size)/(duration) << \" GB/s\" << endl;
   }

assembly code: r10 = bfrptr, r15 = bfrend, rsi = count, rdi = buffer, r13 = k :

$LL5@main:
        mov     r10, rdi
        cmp     rdi, r15
        jae     SHORT $LN4@main
        npad    4
$LL2@main:
        mov     rax, QWORD PTR [r10+24]
        mov     rcx, QWORD PTR [r10+16]
        mov     r8, QWORD PTR [r10+8]
        mov     r9, QWORD PTR [r10]
        popcnt  rdx, rax
        popcnt  rax, rcx
        add     rdx, rax
        popcnt  rax, r8
        add     r10, 32
        add     rdx, rax
        popcnt  rax, r9
        add     rsi, rax
        add     rsi, rdx
        cmp     r10, r15
        jb      SHORT $LL2@main
$LN4@main:
        dec     r13
        jne     SHORT $LL5@main


回答6:

Have you tried passing -funroll-loops -fprefetch-loop-arrays to GCC?

I get the following results with these additional optimizations:

[1829] /tmp/so_25078285 $ cat /proc/cpuinfo |grep CPU|head -n1
model name      : Intel(R) Core(TM) i3-3225 CPU @ 3.30GHz
[1829] /tmp/so_25078285 $ g++ --version|head -n1
g++ (Ubuntu/Linaro 4.7.3-1ubuntu1) 4.7.3

[1829] /tmp/so_25078285 $ g++ -O3 -march=native -std=c++11 test.cpp -o test_o3
[1829] /tmp/so_25078285 $ g++ -O3 -march=native -funroll-loops -fprefetch-loop-arrays -std=c++11     test.cpp -o test_o3_unroll_loops__and__prefetch_loop_arrays

[1829] /tmp/so_25078285 $ ./test_o3 1
unsigned        41959360000     0.595 sec       17.6231 GB/s
uint64_t        41959360000     0.898626 sec    11.6687 GB/s

[1829] /tmp/so_25078285 $ ./test_o3_unroll_loops__and__prefetch_loop_arrays 1
unsigned        41959360000     0.618222 sec    16.9612 GB/s
uint64_t        41959360000     0.407304 sec    25.7443 GB/s


回答7:

Have you tried moving the reduction step outside the loop? Right now you have a data dependency that really isn\'t needed.

Try:

  uint64_t subset_counts[4] = {};
  for( unsigned k = 0; k < 10000; k++){
     // Tight unrolled loop with unsigned
     unsigned i=0;
     while (i < size/8) {
        subset_counts[0] += _mm_popcnt_u64(buffer[i]);
        subset_counts[1] += _mm_popcnt_u64(buffer[i+1]);
        subset_counts[2] += _mm_popcnt_u64(buffer[i+2]);
        subset_counts[3] += _mm_popcnt_u64(buffer[i+3]);
        i += 4;
     }
  }
  count = subset_counts[0] + subset_counts[1] + subset_counts[2] + subset_counts[3];

You also have some weird aliasing going on, that I\'m not sure is conformant to the strict aliasing rules.



回答8:

TL;DR: Use __builtin intrinsics instead.

I was able to make gcc 4.8.4 (and even 4.7.3 on gcc.godbolt.org) generate optimal code for this by using __builtin_popcountll which uses the same assembly instruction, but doesn\'t have that false dependency bug.

I am not 100% sure of my benchmarking code, but objdump output seems to share my views. I use some other tricks (++i vs i++) to make the compiler unroll loop for me without any movl instruction (strange behaviour, I must say).

Results:

Count: 20318230000  Elapsed: 0.411156 seconds   Speed: 25.503118 GB/s

Benchmarking code:

#include <stdint.h>
#include <stddef.h>
#include <time.h>
#include <stdio.h>
#include <stdlib.h>

uint64_t builtin_popcnt(const uint64_t* buf, size_t len){
  uint64_t cnt = 0;
  for(size_t i = 0; i < len; ++i){
    cnt += __builtin_popcountll(buf[i]);
  }
  return cnt;
}

int main(int argc, char** argv){
  if(argc != 2){
    printf(\"Usage: %s <buffer size in MB>\\n\", argv[0]);
    return -1;
  }
  uint64_t size = atol(argv[1]) << 20;
  uint64_t* buffer = (uint64_t*)malloc((size/8)*sizeof(*buffer));

  // Spoil copy-on-write memory allocation on *nix
  for (size_t i = 0; i < (size / 8); i++) {
    buffer[i] = random();
  }
  uint64_t count = 0;
  clock_t tic = clock();
  for(size_t i = 0; i < 10000; ++i){
    count += builtin_popcnt(buffer, size/8);
  }
  clock_t toc = clock();
  printf(\"Count: %lu\\tElapsed: %f seconds\\tSpeed: %f GB/s\\n\", count, (double)(toc - tic) / CLOCKS_PER_SEC, ((10000.0*size)/(((double)(toc - tic)*1e+9) / CLOCKS_PER_SEC)));
  return 0;
}

Compile options:

gcc --std=gnu99 -mpopcnt -O3 -funroll-loops -march=native bench.c -o bench

GCC version:

gcc (Ubuntu 4.8.4-2ubuntu1~14.04.1) 4.8.4

Linux kernel version:

3.19.0-58-generic

CPU information:

processor   : 0
vendor_id   : GenuineIntel
cpu family  : 6
model       : 70
model name  : Intel(R) Core(TM) i7-4870HQ CPU @ 2.50 GHz
stepping    : 1
microcode   : 0xf
cpu MHz     : 2494.226
cache size  : 6144 KB
physical id : 0
siblings    : 1
core id     : 0
cpu cores   : 1
apicid      : 0
initial apicid  : 0
fpu     : yes
fpu_exception   : yes
cpuid level : 13
wp      : yes
flags       : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx rdtscp lm constant_tsc nopl xtopology nonstop_tsc eagerfpu pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm arat pln pts dtherm fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 invpcid xsaveopt
bugs        :
bogomips    : 4988.45
clflush size    : 64
cache_alignment : 64
address sizes   : 36 bits physical, 48 bits virtual
power management:


回答9:

First of all, try to estimate peak performance - examine https://www.intel.com/content/dam/www/public/us/en/documents/manuals/64-ia-32-architectures-optimization-manual.pdf, in particular, Appendix C.

In your case, it\'s table C-10 that shows POPCNT instruction has latency = 3 clocks and throughput = 1 clock. Throughput shows your maximal rate in clocks (multiply by core frequency and 8 bytes in case of popcnt64 to get your best possible bandwidth number).

Now examine what compiler did and sum up throughputs of all other instructions in the loop. This will give best possible estimate for generated code.

At last, look at data dependencies between instructions in the loop as they will force latency-large delay instead of throughput - so split instructions of single iteration on data flow chains and calculate latency across them then naively pick up maximal from them. it will give rough estimate taking into account data flow dependencies.

However, in your case, just writing code the right way would eliminate all these complexities. Instead of accumulating to the same count variable, just accumulate to different ones (like count0, count1, ... count8) and sum them up at the end. Or even create an array of counts[8] and accumulate to its elements - perhaps, it will be vectorized even and you will get much better throughput.

P.S. and never run benchmark for a second, first warm up the core then run loop for at least 10 seconds or better 100 seconds. otherwise, you will test power management firmware and DVFS implementation in hardware :)

P.P.S. I heard endless debates on how much time should benchmark really run. Most smartest folks are even asking why 10 seconds not 11 or 12. I should admit this is funny in theory. In practice, you just go and run benchmark hundred times in a row and record deviations. That IS funny. Most people do change source and run bench after that exactly ONCE to capture new performance record. Do the right things right.

Not convinced still? Just use above C-version of benchmark by assp1r1n3 (https://stackoverflow.com/a/37026212/9706746) and try 100 instead of 10000 in retry loop.

My 7960X shows, with RETRY=100:

Count: 203182300 Elapsed: 0.008385 seconds Speed: 12.505379 GB/s

Count: 203182300 Elapsed: 0.011063 seconds Speed: 9.478225 GB/s

Count: 203182300 Elapsed: 0.011188 seconds Speed: 9.372327 GB/s

Count: 203182300 Elapsed: 0.010393 seconds Speed: 10.089252 GB/s

Count: 203182300 Elapsed: 0.009076 seconds Speed: 11.553283 GB/s

with RETRY=10000:

Count: 20318230000 Elapsed: 0.661791 seconds Speed: 15.844519 GB/s

Count: 20318230000 Elapsed: 0.665422 seconds Speed: 15.758060 GB/s

Count: 20318230000 Elapsed: 0.660983 seconds Speed: 15.863888 GB/s

Count: 20318230000 Elapsed: 0.665337 seconds Speed: 15.760073 GB/s

Count: 20318230000 Elapsed: 0.662138 seconds Speed: 15.836215 GB/s

P.P.P.S. Finally, on \"accepted answer\" and other mistery ;-)

Let\'s use assp1r1n3\'s answer - he has 2.5Ghz core. POPCNT has 1 clock throuhgput, his code is using 64-bit popcnt. So math is 2.5Ghz * 1 clock * 8 bytes = 20 GB/s for his setup. He is seeing 25Gb/s, perhaps due to turbo boost to around 3Ghz.

Thus go to ark.intel.com and look for i7-4870HQ: https://ark.intel.com/products/83504/Intel-Core-i7-4870HQ-Processor-6M-Cache-up-to-3-70-GHz-?q=i7-4870HQ

That core could run up to 3.7Ghz and real maximal rate is 29.6 GB/s for his hardware. So where is another 4GB/s? Perhaps, it\'s spent on loop logic and other surrounding code within each iteration.

Now where is this false dependency? hardware runs at almost peak rate. Maybe my math is bad, it happens sometimes :)

P.P.P.P.P.S. Still people suggesting HW errata is culprit, so I follow suggestion and created inline asm example, see below.

On my 7960X, first version (with single output to cnt0) runs at 11MB/s, second version (with output to cnt0, cnt1, cnt2 and cnt3) runs at 33MB/s. And one could say - voila! it\'s output dependency.

OK, maybe, the point I made is that it does not make sense to write code like this and it\'s not output dependency problem but dumb code generation. We are not testing hardware, we are writing code to unleash maximal performance. You could expect that HW OOO should rename and hide those \"output-dependencies\" but, gash, just do the right things right and you will never face any mystery.

uint64_t builtin_popcnt1a(const uint64_t* buf, size_t len) 
{
    uint64_t cnt0, cnt1, cnt2, cnt3;
    cnt0 = cnt1 = cnt2 = cnt3 = 0;
    uint64_t val = buf[0];
    #if 0
        __asm__ __volatile__ (
            \"1:\\n\\t\"
            \"popcnt %2, %1\\n\\t\"
            \"popcnt %2, %1\\n\\t\"
            \"popcnt %2, %1\\n\\t\"
            \"popcnt %2, %1\\n\\t\"
            \"subq $4, %0\\n\\t\"
            \"jnz 1b\\n\\t\"
        : \"+q\" (len), \"=q\" (cnt0)
        : \"q\" (val)
        :
        );
    #else
        __asm__ __volatile__ (
            \"1:\\n\\t\"
            \"popcnt %5, %1\\n\\t\"
            \"popcnt %5, %2\\n\\t\"
            \"popcnt %5, %3\\n\\t\"
            \"popcnt %5, %4\\n\\t\"
            \"subq $4, %0\\n\\t\"
            \"jnz 1b\\n\\t\"
        : \"+q\" (len), \"=q\" (cnt0), \"=q\" (cnt1), \"=q\" (cnt2), \"=q\" (cnt3)
        : \"q\" (val)
        :
        );
    #endif
    return cnt0;
}