x86-64: strlen-evex performance performance degradation compared to strlen-avx2

[abush wang]

Actually, I was handling performance issue from  libmicro in our distro OS.
I found that the performance degradation of localtime_r benchmark from
libmicro is blame to strlen.
So I abstracted this test case.

On Sat, Apr 27, 2024 at 12:54 AM Sunil Pandey <skpgkp2@gmail.com> wrote:

>
>
> On Fri, Apr 26, 2024 at 6:30 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>
>> On Thu, Apr 25, 2024 at 9:03 PM abush wang <abushwangs@gmail.com> wrote:
>> >
>> > Hi, H.J.
>> > When I test glibc performance between 2.28 and 2.38,
>> > I found there is a performance degradation about strlen.
>> > In fact, this difference comes from __strlen_avx2 and __strlen_evex
>> >
>> > ```
>> > 2.28
>> > __strlen_avx2 () at ../sysdeps/x86_64/multiarch/strlen-avx2.S:42
>> > 42 ENTRY (STRLEN)
>> >
>> >
>> > 2.38
>> > __strlen_evex () at ../sysdeps/x86_64/multiarch/strlen-evex.S:79
>> > 79 ENTRY_P2ALIGN (STRLEN, 6)
>> > ```
>> >
>> > This is my test:
>> > ```
>> > #include <stdio.h>
>> > #include <stdlib.h>
>> > #include <stdint.h>
>> > #include <string.h>
>> >
>> > #define MAX_STRINGS 100
>> >
>> > uint64_t rdtsc() {
>> >     uint32_t lo, hi;
>> >     __asm__ __volatile__ (
>> >         "rdtsc" : "=a"(lo), "=d"(hi)
>> >     );
>> >     return ((uint64_t)hi << 32) | lo;
>> > }
>> >
>> > int main(int argc, char *argv[]) {
>> >     char *input_str[MAX_STRINGS];
>> >     size_t lengths[MAX_STRINGS];
>> >     int num_strings = 0; // Number of input strings
>> >     uint64_t start_cycles, end_cycles;
>> >
>> >     // Parse command line arguments and store pointers in input_str
>> array
>> >     for (int i = 1; i < argc && num_strings < MAX_STRINGS; ++i) {
>> >         input_str[num_strings] = argv[i];
>> >         num_strings++;
>> >     }
>> >
>> >     // Measure the strlen operation for each string
>> >     start_cycles = rdtsc();
>> >     for (int i = 0; i < num_strings; ++i) {
>> >         lengths[i] = strlen(input_str[i]);
>> >     }
>> >     end_cycles = rdtsc();
>> >
>> >     unsigned long long total_cycle = end_cycles - start_cycles;
>> >     unsigned long long av_cycle = total_cycle / num_strings;
>> >     // Print the total cycles taken for the strlen operations
>> >     printf("Total cycles: %llu av cycle: %llu \n", total_cycle,
>> av_cycle);
>> >
>> >     // Print the recorded lengths
>> >     printf("Lengths of the input strings:\n");
>> >     for (int i = 0; i < num_strings; ++i) {
>> >         printf("String %d length: %zu\n", i, lengths[i]);
>> >     }
>> >
>> >     return 0;
>> > }
>> > ```
>> >
>> > This is result
>> > ```
>> > 2.28
>> > ./strlen_test str1 str2 str3 str4 str5
>> > Total cycles: 1468 av cycle: 293
>> > Lengths of the input strings:
>> > String 0 length: 4
>> > String 1 length: 4
>> > String 2 length: 4
>> > String 3 length: 4
>> > String 4 length: 4
>> >
>> > 2.38
>> > ./strlen_test str1 str2 str3 str4 str5
>> > Total cycles: 1814 av cycle: 362
>> > Lengths of the input strings:
>> > String 0 length: 4
>> > String 1 length: 4
>> > String 2 length: 4
>> > String 3 length: 4
>> > String 4 length: 4
>> > ```
>> >
>> > Thanks,
>> > abush
>>
>
> I'm not sure how you are measuring the performance of strlen function.
> Are you making performance conclusion based on these 2 runs?
>
> 2.28
> Total cycles: 1468 av cycle: 293
>
> 2.38
> Total cycles: 1814 av cycle: 362
>
> Please use glibc microbenchmark to see if you can reproduce perf drop.
>
>
>>
>> Which processors did you use?  Sunil, Noah, can we reproduce it?
>>
>> --
>> H.J.
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://sourceware.org/pipermail/libc-alpha/attachments/20240428/0ce08bfe/attachment.htm>