memcount_avx2.c

size_t memcount_avx2(const void *s, int c, size_t n) {    
    __m256i cv = _mm256_set1_epi8(c), zv = _mm256_setzero_si256(), sum = zv, acr0,acr1,acr2,acr3;
    const char *p,*pe;						    											   
	for(p = s; p != (char *)s+(n- (n % (252*32)));) { 
	  for(acr0 = acr1 = acr2 = acr3 = zv,pe = p+252*32; p != pe; p += 128) { 
		acr0 = _mm256_add_epi8(acr0, _mm256_cmpeq_epi8(cv, _mm256_lddqu_si256((const __m256i *)p))); 
		acr1 = _mm256_add_epi8(acr1, _mm256_cmpeq_epi8(cv, _mm256_lddqu_si256((const __m256i *)(p+32)))); 
		acr2 = _mm256_add_epi8(acr2, _mm256_cmpeq_epi8(cv, _mm256_lddqu_si256((const __m256i *)(p+64)))); 
		acr3 = _mm256_add_epi8(acr3, _mm256_cmpeq_epi8(cv, _mm256_lddqu_si256((const __m256i *)(p+96)))); __builtin_prefetch(p+1024);
	  }
      sum = _mm256_add_epi64(sum, _mm256_sad_epu8(_mm256_sub_epi8(zv, acr0), zv));
      sum = _mm256_add_epi64(sum, _mm256_sad_epu8(_mm256_sub_epi8(zv, acr1), zv));
      sum = _mm256_add_epi64(sum, _mm256_sad_epu8(_mm256_sub_epi8(zv, acr2), zv));
      sum = _mm256_add_epi64(sum, _mm256_sad_epu8(_mm256_sub_epi8(zv, acr3), zv));
    }	
    for(acr0=zv; p+32 < (char *)s + n; p += 32)  
      acr0 = _mm256_add_epi8(acr0, _mm256_cmpeq_epi8(cv, _mm256_lddqu_si256((const __m256i *)p))); 
    sum = _mm256_add_epi64(sum, _mm256_sad_epu8(_mm256_sub_epi8(zv, acr0), zv));
    size_t count = _mm256_extract_epi64(sum, 0) + _mm256_extract_epi64(sum, 1) + _mm256_extract_epi64(sum, 2) + _mm256_extract_epi64(sum, 3);   
    while(p != (char *)s + n) count += *p++ == c;
    return count;
}

Hey @powturbo have you checked whether acr's actually get full/reach 127 in this scenario, from all my testing it only got up to 3f/63, could you only use acr0 and acr1 like

acr0 = _mm256_sub_epi8(acr0
acr0 = _mm256_sub_epi8(acr0
acr1 = _mm256_sub_epi8(acr1
acr1 = _mm256_sub_epi8(acr1...

and then only do 2 sad_epu8 calls.

I'm unable to verify if it's faster because even when the data is in cache I'm not hitting 100% core utilization so it waits for memory/cache

I'm using your variant from SO https://stackoverflow.com/a/57929966/23017301

There's a possibility I'm missing something which is why im asking you :P

Okay some further investigation, unrolling the loop to 3 or 5 is better than 4 because then you can set the limit to 255 * 32 and still use a single acr

v32_i8: ff ff ff ff | ff ff ff ff | ff ff ff ff | ff ff ff ff | ff ff ff ff | ff ff ff ff | ff ff ff ff | ff ff ff ff <- acr0
v4_i64: 14280 14280 14280 14280 | 37c8 37c8 37c8 37c8 <- sum
v32_i8: 1 1 1 1 | 1 1 1 1 | 1 1 1 1 | 1 1 1 1 | 1 1 1 1 | 1 1 1 1 | 1 1 1 1 | 1 1 1 1 <- back to the outer loop