Performance on Apple M1

[Mike Taghavi]

Hello, I have been testing the code from this blog on Apple M1 using sse2neon library. However I have encountered a problem with low performance. I am not exactly sure why the result is opposite of the original post on x86_64 in terms of performance as I am a beginner and learning SIMD.

The following program loads an text file with 1.6 Gb size and counts the commas.

#include <stdio.h>    
#include <assert.h>
#include <strings.h>
#include <stdlib.h>
#include <sys/time.h>
#include "simd.h"

// count_chars_8 is copied from Josh Weinstein's Blog
size_t
count_chars_8(const char* data, size_t size, const char ch)
{
	size_t total = 0;
	while (size) {
		if (*data == ch)
			total += 1;
		data += 1;
		size -= 1;
	} 
	return total;
};


// count_chars_128 is copied from Josh Weinstein's Blog
size_t
count_chars_128(const char* data, size_t size, const char ch)
{
	size_t total = 0;
	assert(size % 16 == 0);
	__m128i tocmp =  _mm_set1_epi8(ch);
	while (size) {
		int mask = 0;
		__m128i chunk = _mm_load_si128 ((__m128i const*)data);
		__m128i results =  _mm_cmpeq_epi8(chunk, tocmp);
		mask = _mm_movemask_epi8(results);
		total += _mm_popcnt_u32(mask);
		data += 16;
		size -= 16;
	}
	return total;
};

void
simd(char *string, size_t ln) 
{
    struct timespec start, end;
    clock_gettime(CLOCK_MONOTONIC_RAW, &start);    
    size_t result = count_chars_128(string, ln, ',');
    clock_gettime(CLOCK_MONOTONIC_RAW, &end);
    uint64_t delta_us = (end.tv_sec - start.tv_sec) * 1000000 + (end.tv_nsec - start.tv_nsec) / 1000;
    printf("[x] SIMD   took %llu u/s\ncount: %d\n", delta_us, result);
};

void
normal(char *string) 
{
    struct timespec start, end;
    clock_gettime(CLOCK_MONOTONIC_RAW, &start);    
    size_t result = count_chars_8(string, strlen(string), ',');
    clock_gettime(CLOCK_MONOTONIC_RAW, &end);
    uint64_t delta_us = (end.tv_sec - start.tv_sec) * 1000000 + (end.tv_nsec - start.tv_nsec) / 1000;
    printf("[x] NORMAL took %llu u/s\ncount: %d\n", delta_us, result);
};


int
main() 
{
    FILE *f = fopen("/tmp/big.txt", "rb"); // 1.6 Gb file containing repeated "tes,tes,tes, ..." string
    fseek(f, 0, SEEK_END);
    long fsize = ftell(f);
    fseek(f, 0, SEEK_SET);

    char *string = malloc(fsize + 1);
    fread(string, fsize, 1, f);
    fclose(f);

    string[fsize] = 0;

    puts("[+] File loaded");
    
    normal(string);      
    simd(string, fsize); // <- usually 5-6 times slower than non-simd version

    free(string);    

    return 0;        
};

Here is the result I get:

MBPMIL:simd mitghi$ ./main
[+] File loaded
[x] NORMAL took 2165755 u/s
count: 500000000
[x] SIMD   took 7884030 u/s
count: 500000000

The SIMD version is most of the times 5-6 times slower than the non-SIMD version which is the opposite of benchmark on x86_64.

Could you please provide some insight and assistance?

Thanks

[Jim Huang]

Since SSE2NEON is just a translator from SSE intrinsics to NEON counterpart, it might not be worthy to take SSE tweaked routines for IO-bound scenario such as count_chars_128. Instead, you should check some SWAR/NEON implementations. For example, check this: https://github.com/magurosan/strlen_neon

[Mike Taghavi]

@jserv Thank you very much for the quick reply. The memchr implementation in strlen_neon has almost everything I was looking for.