//Note: it takes size and offset in units of byte
static inline int compute_ham_similarity_64(unsigned short* ref, unsigned short* circ_array,
int size){
const uint8_t* ref_c=(uint8_t*) ref;
const uint8_t* circ_c=(uint8_t*) circ_array;
register uint8x16_t a,b;
register uint8x16_t c,d,temp;
register uint16x8_t acc;
register uint i=0,count=0;
int j=0;
int shift=size&0xF;
for(i=0;i<=size-16; i+=16){
j++;
a=vld1q_u8(&ref_c[i]);
b=vld1q_u8(&circ_c[i]);
c=veorq_u8(a,b);
acc=vaddq_u16(acc,vpaddlq_u8(vcntq_u8(c)));
}
count=setbits(acc);
a=vld1q_u8(&ref_c[i]);
b=vld1q_u8(&circ_c[i]);
c=veorq_u8(a,b);
c=vcntq_u8(c);
for(i=0;i<shift;i++){
count=count+vgetq_lane_u8 (c,i);
}
return size*8-count;
}
uint32x2_t FORCE_INLINE popcnt_neon_qreg(const uint8x16_t reg) {
const uint8x16_t pcnt = vcntq_u8(reg);
const uint16x8_t t0 = vpaddlq_u8(pcnt);
const uint32x4_t t1 = vpaddlq_u16(t0);
const uint32x2_t t2 = vadd_u32(vget_low_u32(t1), vget_high_u32(t1));
return t2;
}
uint64_t popcnt_neon_vcnt(const uint8_t* data, const size_t size)
{
const size_t chunk_size = 16 * 4 * 2;
uint8_t* ptr = const_cast<uint8_t*>(data);
const size_t n = size / chunk_size;
const size_t k = size % chunk_size;
uint32x4_t sum = vcombine_u32(vcreate_u32(0), vcreate_u32(0));
for (size_t i=0; i < n; i++, ptr += chunk_size) {
uint8x16x4_t input0 = vld4q_u8(ptr + 0 * 16 * 4);
uint8x16x4_t input1 = vld4q_u8(ptr + 1 * 16 * 4);
uint8x16_t t0 = vcntq_u8(input0.val[0]);
t0 = vaddq_u8(t0, vcntq_u8(input0.val[1]));
t0 = vaddq_u8(t0, vcntq_u8(input0.val[2]));
t0 = vaddq_u8(t0, vcntq_u8(input0.val[3]));
t0 = vaddq_u8(t0, vcntq_u8(input1.val[0]));
t0 = vaddq_u8(t0, vcntq_u8(input1.val[1]));
t0 = vaddq_u8(t0, vcntq_u8(input1.val[2]));
t0 = vaddq_u8(t0, vcntq_u8(input1.val[3]));
const uint16x8_t t1 = vpaddlq_u8(t0);
sum = vpadalq_u16(sum, t1);
}
uint32_t scalar = 0;
uint32_t tmp[4];
vst1q_u32(tmp, sum);
for (int i=0; i < 4; i++) {
scalar += tmp[i];
}
for (size_t j=0; j < k; j++) {
scalar += lookup8bit[ptr[j]];
}
return scalar;
}
size_t mempopcnt(const void *s, size_t len)
{
uint8x16_t v_0;
uint8x16_t c;
uint32x4_t v_sum;
uint32x2_t v_tsum;
unsigned char *p;
size_t r;
unsigned shift;
prefetch(s);
// TODO: do this in 64 bit? the mem model seems more that way...
v_0 = (uint8x16_t){0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
v_sum = (uint32x4_t)v_0;
p = (unsigned char *)ALIGN_DOWN(s, SOVUCQ);
shift = ALIGN_DOWN_DIFF(s, SOVUCQ);
c = *(const uint8x16_t *)p;
if(HOST_IS_BIGENDIAN)
c = neon_simple_alignq(v_0, c, SOVUCQ - shift);
else
c = neon_simple_alignq(c, v_0, shift);
if(len >= SOVUCQ || len + shift >= SOVUCQ)
{
p += SOVUCQ;
len -= SOVUCQ - shift;
v_sum = vpadalq_u16(v_sum, vpaddlq_u8(vcntq_u8(c)));
while(len >= SOVUCQ * 2)
{
uint8x16_t v_sumb = v_0;
r = len / (SOVUCQ * 2);
r = r > 15 ? 15 : r;
len -= r * SOVUCQ * 2;
/*
* NEON has a vector popcnt instruction, so no compression.
* We trust the speed given in the handbook (adding more
* instructions would not make it faster), 1-2 cycles.
*/
for(; r; r--, p += SOVUCQ * 2) {
c = *(const uint8x16_t *)p;
v_sumb = vaddq_u8(v_sumb, vcntq_u8(c));
c = *((const uint8x16_t *)(p + SOVUCQ));
v_sumb = vaddq_u8(v_sumb, vcntq_u8(c));
}
v_sum = vpadalq_u16(v_sum, vpaddlq_u8(v_sumb));
}
if(len >= SOVUCQ) {
c = *(const uint8x16_t *)p;
p += SOVUCQ;
v_sum = vpadalq_u16(v_sum, vpaddlq_u8(vcntq_u8(c)));
len -= SOVUCQ;
}
if(len)
c = *(const uint8x16_t *)p;
}
if(len) {
if(HOST_IS_BIGENDIAN)
c = neon_simple_alignq(c, v_0, SOVUCQ - len);
else
c = neon_simple_alignq(v_0, c, len);
v_sum = vpadalq_u16(v_sum, vpaddlq_u8(vcntq_u8(c)));
}
v_tsum = vpadd_u32(vget_high_u32(v_sum), vget_low_u32(v_sum));
v_tsum = vpadd_u32(v_tsum, v_tsum);
return vget_lane_u32(v_tsum, 0);
}
inline ResultType operator()(Iterator1 a, Iterator2 b, size_t size) const
{
ResultType result = 0;
#if (defined __GNUC__ || defined __clang__) && defined USE_SSE
#ifdef __ARM_NEON__
{
uint32x4_t bits = vmovq_n_u32(0);
for (size_t i = 0; i < size; i += 16) {
uint8x16_t A_vec = vld1q_u8 (a + i);
uint8x16_t B_vec = vld1q_u8 (b + i);
uint8x16_t AxorB = veorq_u8 (A_vec, B_vec);
uint8x16_t bitsSet = vcntq_u8 (AxorB);
uint16x8_t bitSet8 = vpaddlq_u8 (bitsSet);
uint32x4_t bitSet4 = vpaddlq_u16 (bitSet8);
bits = vaddq_u32(bits, bitSet4);
}
uint64x2_t bitSet2 = vpaddlq_u32 (bits);
result = vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),0);
result += vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),2);
}
#else
{
//for portability just use unsigned long -- and use the __builtin_popcountll (see docs for __builtin_popcountll)
typedef unsigned long long pop_t;
const size_t modulo = size % sizeof(pop_t);
const pop_t* a2 = reinterpret_cast<const pop_t*> (a);
const pop_t* b2 = reinterpret_cast<const pop_t*> (b);
const pop_t* a2_end = a2 + (size / sizeof(pop_t));
for (; a2 != a2_end; ++a2, ++b2) result += __builtin_popcountll((*a2) ^ (*b2));
if (modulo) {
//in the case where size is not dividable by sizeof(pop_t)
//need to mask off the bits at the end
pop_t a_final = 0, b_final = 0;
memcpy(&a_final, a2, modulo);
memcpy(&b_final, b2, modulo);
result += __builtin_popcountll(a_final ^ b_final);
}
}
#endif //NEON
return result;
#endif
#ifdef PLATFORM_64_BIT
if(size%64 == 0)
{
const uint64_t* pa = reinterpret_cast<const uint64_t*>(a);
const uint64_t* pb = reinterpret_cast<const uint64_t*>(b);
size /= (sizeof(uint64_t)/sizeof(unsigned char));
for(size_t i = 0; i < size; ++i, ++pa, ++pb ) {
result += popcnt64(*pa ^ *pb);
}
}
else
{
const uint32_t* pa = reinterpret_cast<const uint32_t*>(a);
const uint32_t* pb = reinterpret_cast<const uint32_t*>(b);
size /= (sizeof(uint32_t)/sizeof(unsigned char));
for(size_t i = 0; i < size; ++i, ++pa, ++pb ) {
result += popcnt32(*pa ^ *pb);
}
}
#else
const uint32_t* pa = reinterpret_cast<const uint32_t*>(a);
const uint32_t* pb = reinterpret_cast<const uint32_t*>(b);
size /= (sizeof(uint32_t)/sizeof(unsigned char));
for(size_t i = 0; i < size; ++i, ++pa, ++pb ) {
result += popcnt32(*pa ^ *pb);
}
#endif
return result;
}
