/// The thread entry for the calculation of LD measure
static void Entry_LD_Matrix(PdThread Thread, int ThreadIndex, void *Param)
{
double *base = (double*)Param;
IdMatTri I = Array_Thread_MatIdx[ThreadIndex];
C_UInt8 *pGeno = &PackedGeno[0];
for (C_Int64 n=Array_Thread_MatCnt[ThreadIndex]; n > 0; n--)
{
const int i = I.Row(), j = I.Column();
double &p1 = base[i*NumSNP + j];
double &p2 = base[j*NumSNP + i];
C_UInt8 *s1 = pGeno + i*nPackedSamp;
C_UInt8 *s2 = pGeno + j*nPackedSamp;
switch (LD_Method)
{
case 1:
p1 = p2 = PairComposite(s1, s2); break;
case 2:
p1 = p2 = PairR(s1, s2); break;
case 3:
p1 = p2 = PairDPrime(s1, s2); break;
case 4:
p1 = p2 = PairCorr(s1, s2); break;
case 5:
p1 = p2 = PairCov(s1, s2); break;
default:
p1 = p2 = R_NaN;
}
++ I;
}
}
/// Compute IBD estimator in KING-h**o
static void _Do_KING_Homo_Compute(int ThreadIndex, long Start, long SNP_Cnt, void* Param)
{
long Cnt = IBS_Thread_MatCnt[ThreadIndex];
IdMatTri I = IBS_Thread_MatIdx[ThreadIndex];
TKINGHomoFlag *p = ((TKINGHomoFlag*)Param) + I.Offset();
long _PackSNPLen = (SNP_Cnt / 4) + (SNP_Cnt % 4 ? 1 : 0);
for (; Cnt > 0; Cnt--, ++I, p++)
{
UInt8 *p1 = GenoPacked.get() + I.Row()*_PackSNPLen;
UInt8 *p2 = GenoPacked.get() + I.Column()*_PackSNPLen;
for (long k=0; k < _PackSNPLen; k++, p1++, p2++)
{
size_t t = (size_t(*p1) << 8) | (*p2);
p->IBS0 += IBS0_Num_SNP[t];
p->SumSq += Gen_KING_SqDiff[t];
UInt8 flag = Gen_Both_Valid[t];
if (flag & 0x01)
{
double f = GenoAlleleFreq.get()[4*k + 0];
p->SumAFreq += f; p->SumAFreq2 += f*f;
}
if (flag & 0x02)
{
double f = GenoAlleleFreq.get()[4*k + 1];
p->SumAFreq += f; p->SumAFreq2 += f*f;
}
if (flag & 0x04)
{
double f = GenoAlleleFreq.get()[4*k + 2];
p->SumAFreq += f; p->SumAFreq2 += f*f;
}
if (flag & 0x08)
{
double f = GenoAlleleFreq.get()[4*k + 3];
p->SumAFreq += f; p->SumAFreq2 += f*f;
}
}
}
}
/// Compute the pairwise IBS matrix
static void _Do_IBS_Compute(int ThreadIndex, long Start, long SNP_Cnt, void* Param)
{
long Cnt = IBS_Thread_MatCnt[ThreadIndex];
IdMatTri I = IBS_Thread_MatIdx[ThreadIndex];
TIBS_Flag *p = ((TIBS_Flag*)Param) + I.Offset();
long _PackSNPLen = (SNP_Cnt / 4) + (SNP_Cnt % 4 ? 1 : 0);
for (; Cnt > 0; Cnt--, ++I, p++)
{
UInt8 *p1 = GenoPacked.get() + I.Row()*_PackSNPLen;
UInt8 *p2 = GenoPacked.get() + I.Column()*_PackSNPLen;
for (long k=_PackSNPLen; k > 0; k--, p1++, p2++)
{
size_t t = (size_t(*p1) << 8) | (*p2);
p->IBS0 += IBS0_Num_SNP[t];
p->IBS1 += IBS1_Num_SNP[t];
p->IBS2 += IBS2_Num_SNP[t];
}
}
}
/// Compute IBD estimator in KING-robust
static void _Do_KING_Robust_Compute(int ThreadIndex, long Start, long SNP_Cnt, void* Param)
{
long Cnt = IBS_Thread_MatCnt[ThreadIndex];
IdMatTri I = IBS_Thread_MatIdx[ThreadIndex];
TKINGRobustFlag *p = ((TKINGRobustFlag*)Param) + I.Offset();
long _PackSNPLen = (SNP_Cnt / 4) + (SNP_Cnt % 4 ? 1 : 0);
for (; Cnt > 0; Cnt--, ++I, p++)
{
UInt8 *p1 = GenoPacked.get() + I.Row()*_PackSNPLen;
UInt8 *p2 = GenoPacked.get() + I.Column()*_PackSNPLen;
for (long k=0; k < _PackSNPLen; k++, p1++, p2++)
{
size_t t = (size_t(*p1) << 8) | (*p2);
p->IBS0 += IBS0_Num_SNP[t];
p->nLoci += Gen_KING_Num_Loci[t];
p->SumSq += Gen_KING_SqDiff[t];
p->N1_Aa += Gen_KING_N1_Aa[t];
p->N2_Aa += Gen_KING_N2_Aa[t];
}
}
}
/// Compute the covariate matrix
static void _Do_Diss_Compute(int ThreadIndex, long Start, long SNP_Cnt, void* Param)
{
long Cnt = IBS_Thread_MatCnt[ThreadIndex];
IdMatTri I = IBS_Thread_MatIdx[ThreadIndex];
TDissflag *p = ((TDissflag*)Param) + I.Offset();
long _PackSNPLen = (SNP_Cnt / 4) + (SNP_Cnt % 4 ? 1 : 0);
for (; Cnt > 0; Cnt--, ++I, p++)
{
UInt8 *p1 = GenoPacked.get() + I.Row()*_PackSNPLen;
UInt8 *p2 = GenoPacked.get() + I.Column()*_PackSNPLen;
for (long k=0; k < _PackSNPLen; k++, p1++, p2++)
{
size_t t = (size_t(*p1) << 8) | (*p2);
p->SumGeno += Gen_Dist_SNP[t];
UInt8 flag = Gen_Both_Valid[t];
if (flag & 0x01) p->SumAFreq += GenoAlleleFreq.get()[4*k];
if (flag & 0x02) p->SumAFreq += GenoAlleleFreq.get()[4*k+1];
if (flag & 0x04) p->SumAFreq += GenoAlleleFreq.get()[4*k+2];
if (flag & 0x08) p->SumAFreq += GenoAlleleFreq.get()[4*k+3];
}
}
}
void thread_ibs_num(size_t i, size_t n)
{
const size_t npack = nBlock >> 3;
const size_t npack2 = npack * 2;
C_UInt8 *Base = Geno1b.Get();
IdMatTri I = Array_Thread_MatIdx[i];
C_Int64 N = Array_Thread_MatCnt[i];
TS_KINGHomo *p = ptrKING + I.Offset();
for (; N > 0; N--, ++I, p++)
{
C_UInt8 *p1 = Base + I.Row() * npack2;
C_UInt8 *p2 = Base + I.Column() * npack2;
double *pAF = AF_1_AF.Get();
double *pAF2 = AF_1_AF_2.Get();
ssize_t m = npack;
#if defined(COREARRAY_SIMD_SSE2)
{
POPCNT_SSE2_HEAD
__m128i ibs0_sum, sumsq_sum;
ibs0_sum = sumsq_sum = _mm_setzero_si128();
__m128d sq_sum, sq_sum2;
sq_sum = sq_sum2 = _mm_setzero_pd();
for (; m > 0; m-=16)
{
__m128i g1_1 = _mm_load_si128((__m128i*)p1);
__m128i g1_2 = _mm_load_si128((__m128i*)(p1 + npack));
__m128i g2_1 = _mm_load_si128((__m128i*)p2);
__m128i g2_2 = _mm_load_si128((__m128i*)(p2 + npack));
p1 += 16; p2 += 16;
__m128i mask = (g1_1 | ~g1_2) & (g2_1 | ~g2_2);
__m128i ibs0 = (~((g1_1 ^ ~g2_1) | (g1_2 ^ ~g2_2))) & mask;
__m128i het = ((g1_1 ^ g1_2) ^ (g2_1 ^ g2_2)) & mask;
POPCNT_SSE2_RUN(ibs0)
ibs0_sum = _mm_add_epi32(ibs0_sum, ibs0);
POPCNT_SSE2_RUN(het)
sumsq_sum = _mm_add_epi32(_mm_add_epi32(sumsq_sum, het),
_mm_slli_epi32(ibs0, 2));
C_UInt64 m1 = _mm_cvtsi128_si64(mask);
C_UInt64 m2 = _mm_cvtsi128_si64(_mm_shuffle_epi32(mask,
_MM_SHUFFLE(1,0,3,2)));
for (size_t k=32; k > 0; k--)
{
switch (m1 & 0x03)
{
case 3:
sq_sum = _mm_add_pd(sq_sum, _mm_load_pd(pAF));
sq_sum2 = _mm_add_pd(sq_sum2, _mm_load_pd(pAF2));
break;
case 1:
sq_sum = _mm_add_pd(sq_sum, _mm_set_pd(0, pAF[0]));
sq_sum2 = _mm_add_pd(sq_sum2, _mm_set_pd(0, pAF2[0]));
break;
case 2:
sq_sum = _mm_add_pd(sq_sum, _mm_set_pd(pAF[1], 0));
sq_sum2 = _mm_add_pd(sq_sum2, _mm_set_pd(pAF2[1], 0));
break;
}
pAF += 2; pAF2 += 2; m1 >>= 2;
}
for (size_t k=32; k > 0; k--)
{
switch (m2 & 0x03)
{
case 3:
sq_sum = _mm_add_pd(sq_sum, _mm_load_pd(pAF));
sq_sum2 = _mm_add_pd(sq_sum2, _mm_load_pd(pAF2));
break;
case 1:
sq_sum = _mm_add_pd(sq_sum, _mm_set_pd(0, pAF[0]));
sq_sum2 = _mm_add_pd(sq_sum2, _mm_set_pd(0, pAF2[0]));
break;
case 2:
sq_sum = _mm_add_pd(sq_sum, _mm_set_pd(pAF[1], 0));
sq_sum2 = _mm_add_pd(sq_sum2, _mm_set_pd(pAF2[1], 0));
break;
}
pAF += 2; pAF2 += 2; m2 >>= 2;
}
}
p->IBS0 += vec_sum_i32(ibs0_sum);
p->SumSq += vec_sum_i32(sumsq_sum);
p->SumAFreq += vec_sum_f64(sq_sum);
p->SumAFreq2 += vec_sum_f64(sq_sum2);
}
#else
for (; m > 0; m-=8)
{
C_UInt64 g1_1 = *((C_UInt64*)p1);
C_UInt64 g1_2 = *((C_UInt64*)(p1 + npack));
C_UInt64 g2_1 = *((C_UInt64*)p2);
C_UInt64 g2_2 = *((C_UInt64*)(p2 + npack));
p1 += 8; p2 += 8;
C_UInt64 mask = (g1_1 | ~g1_2) & (g2_1 | ~g2_2);
C_UInt64 ibs0 = (~((g1_1 ^ ~g2_1) | (g1_2 ^ ~g2_2))) & mask;
C_UInt64 het = ((g1_1 ^ g1_2) ^ (g2_1 ^ g2_2)) & mask;
p->IBS0 += POPCNT_U64(ibs0);
p->SumSq += POPCNT_U64(het) + POPCNT_U64(ibs0)*4;
double sum=0, sum2=0;
for (size_t k=64; k > 0; k--)
{
if (mask & 0x01)
{ sum += (*pAF); sum2 += (*pAF2); }
pAF ++; pAF2 ++;
mask >>= 1;
}
p->SumAFreq += sum;
p->SumAFreq2 += sum2;
}
#endif
}
}
