BOOST_FORCEINLINE
__m256i shuffle(__m256i const lower, __m256i const upper)
{
return _mm256_castpd_si256(
_mm256_shuffle_pd( _mm256_castsi256_pd(lower), _mm256_castsi256_pd(upper)
, _MM_SHUFFLE(upper_i1, upper_i0, lower_i1, lower_i0)
)
);
}
int main()
{
__m256d a;
__m256i mask;
double d[4]={1,2,3,4};
a = _mm256_setzero_pd();
mask = _mm256_castpd_si256(a);
#ifdef GMX_SIMD_X86_AVX_GCC_MASKLOAD_BUG
a = _mm256_maskload_pd(d,_mm256_castsi256_pd(mask));
#else
a = _mm256_maskload_pd(d,mask);
#endif
}
void calculate_fma_double (unsigned char * out, double X0, double Y0, double scale, unsigned YSTART, unsigned SX, unsigned SY)
{
__m256d dd = _mm256_set1_pd (scale);
__m256d XX0 = _mm256_set1_pd (X0);
for (unsigned j = YSTART; j < SY; j++) {
__m256d y0 = _mm256_set1_pd (j*scale + Y0);
for (unsigned i = 0; i < SX; i += 4) {
__m128i ind = _mm_setr_epi32 (i, i + 1, i + 2, i + 3);
__m256d x0 = _mm256_fmadd_pd (dd, _mm256_cvtepi32_pd (ind), XX0);
__m256d x = x0;
__m256d y = y0;
__m256i counts = _mm256_setzero_si256 ();
__m256i cmp_mask = _mm256_set1_epi32 (0xFFFFFFFFu);
for (unsigned n = 0; n < 255; n++) {
__m256d x2 = _mm256_mul_pd (x, x);
__m256d y2 = _mm256_mul_pd (y, y);
__m256d abs = _mm256_add_pd (x2, y2);
__m256i cmp = _mm256_castpd_si256 (_mm256_cmp_pd (abs, _mm256_set1_pd (4), 1));
cmp_mask = _mm256_and_si256 (cmp_mask, cmp);
if (_mm256_testz_si256 (cmp_mask, cmp_mask)) {
break;
}
counts = _mm256_sub_epi64 (counts, cmp_mask);
__m256d t = _mm256_add_pd (x, x);
y = _mm256_fmadd_pd (t, y, y0);
x = _mm256_add_pd (_mm256_sub_pd (x2, y2), x0);
}
__m256i result = _mm256_shuffle_epi8 (counts, _mm256_setr_epi8 (0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8));
*(uint32_t*) out = _mm_extract_epi16 (_mm256_extracti128_si256 (result, 0), 0) | (_mm_extract_epi16 (_mm256_extracti128_si256 (result, 1), 0) << 16);
out += 4;
}
}
}
double bst_compute_129_m256_maskstore_root_aligned( void*_bst_obj, double* p, double* q, size_t nn ) {
segments_t* mem = (segments_t*) _bst_obj;
int n, i, r, l_end, j, l_end_pre;
double t, e_tmp;
double* e = mem->e, *w = mem->w;
int* root = mem->r;
__m256d v_tmp;
__m256d v00, v01, v02, v03;
__m256d v10, v11, v12, v13;
__m256d v20, v21, v22, v23;
__m256d v30, v31, v32, v33;
__m256i v_cur_roots;
__m256 v_rootmask0, v_rootmask1;
// initialization
// mem->n = nn;
n = nn; // subtractions with n potentially negative. say hello to all the bugs
int idx1, idx1_root;
int idx2;
int idx3, idx3_root;
int pad_root, pad, pad_r;
idx1 = ((int) mem->e_sz) - 1;
idx1_root = ((int) mem->r_sz);
// the conventio is that iteration i, idx1 points to the first element of line i+1
e[idx1++] = q[n];
// pad contains the padding for row i+1
// for row n it's always 3
pad = 3;
pad_root = 7;
for (i = n-1; i >= 0; --i) {
idx1 -= 2*(n-i)+1 + pad;
idx1_root -= 2*(n-i)+1 + pad_root;
idx2 = idx1 + 1;
e[idx1] = q[i];
w[idx1] = q[i];
for (j = i+1; j < n+1; ++j,++idx2) {
e[idx2] = INFINITY;
w[idx2] = w[idx2-1] + p[j-1] + q[j];
}
idx2 += pad; // padding of line i+1
// idx2 now points to the first element of the next line
idx3 = idx1;
idx3_root = idx1_root;
pad_r = pad;
for (r = i; r < n; ++r) {
pad_r = (pad_r+1)&3; // padding of line r+1
idx1 = idx3;
idx1_root = idx3_root;
l_end = idx2 + (n-r);
// l_end points to the first entry after the current row
e_tmp = e[idx1++];
idx1_root++;
// calculate until a multiple of 8 doubles is left
// 8 = 4 * 2 128-bit vectors
l_end_pre = idx2 + ((n-r)&15);
for( ; (idx2 < l_end_pre) && (idx2 < l_end); ++idx2 ) {
t = e_tmp + e[idx2] + w[idx1];
if (t < e[idx1]) {
e[idx1] = t;
root[idx1_root] = r;
}
idx1++;
idx1_root++;
}
v_tmp = _mm256_set_pd( e_tmp, e_tmp, e_tmp, e_tmp );
// execute the shit for 4 vectors of size 2
v_cur_roots = _mm256_set_epi32(r, r, r, r, r, r, r, r);
for( ; idx2 < l_end; idx2 += 16 ) {
v01 = _mm256_load_pd( &w[idx1 ] );
v11 = _mm256_load_pd( &w[idx1+ 4] );
v21 = _mm256_load_pd( &w[idx1+ 8] );
v31 = _mm256_load_pd( &w[idx1+12] );
v00 = _mm256_load_pd( &e[idx2 ] );
v01 = _mm256_add_pd( v01, v_tmp );
v10 = _mm256_load_pd( &e[idx2+ 4] );
v11 = _mm256_add_pd( v11, v_tmp );
v20 = _mm256_load_pd( &e[idx2+ 8] );
v21 = _mm256_add_pd( v21, v_tmp );
v30 = _mm256_load_pd( &e[idx2+12] );
v31 = _mm256_add_pd( v31, v_tmp );
v01 = _mm256_add_pd( v01, v00 );
v03 = _mm256_load_pd( &e[idx1 ] );
v11 = _mm256_add_pd( v11, v10 );
v13 = _mm256_load_pd( &e[idx1+ 4] );
v21 = _mm256_add_pd( v21, v20 );
v23 = _mm256_load_pd( &e[idx1+ 8] );
v31 = _mm256_add_pd( v31, v30 );
v33 = _mm256_load_pd( &e[idx1+12] );
v02 = _mm256_cmp_pd( v01, v03, _CMP_LT_OQ );
v12 = _mm256_cmp_pd( v11, v13, _CMP_LT_OQ );
v22 = _mm256_cmp_pd( v21, v23, _CMP_LT_OQ );
v32 = _mm256_cmp_pd( v31, v33, _CMP_LT_OQ );
_mm256_maskstore_pd( &e[idx1 ],
_mm256_castpd_si256( v02 ), v01 );
_mm256_maskstore_pd( &e[idx1+ 4],
_mm256_castpd_si256( v12 ), v11 );
v_rootmask0 = _mm256_insertf128_ps(
_mm256_castps128_ps256(
_mm256_cvtpd_ps(v02)),
_mm256_cvtpd_ps(v12) , 1
);
_mm256_maskstore_pd( &e[idx1+ 8],
_mm256_castpd_si256( v22 ), v21 );
_mm256_maskstore_pd( &e[idx1+12],
_mm256_castpd_si256( v32 ), v31 );
v_rootmask1 = _mm256_insertf128_ps(
_mm256_castps128_ps256(
_mm256_cvtpd_ps(v22)),
_mm256_cvtpd_ps(v32) , 1
);
_mm256_maskstore_ps( &root[idx1_root ],
_mm256_castps_si256( v_rootmask0 ),
_mm256_castsi256_ps( v_cur_roots ) );
_mm256_maskstore_ps( &root[idx1_root + 8],
_mm256_castps_si256( v_rootmask1 ),
_mm256_castsi256_ps( v_cur_roots ) );
idx1 += 16;
idx1_root += 16;
}
idx2 += pad_r;
idx3++;
idx3_root++;
}
pad = (pad -1)&3;
pad_root = (pad_root-1)&7;
}
// the index of the last item of the first row is ((n/4)+1)*4-1, due to the padding
// if n is even, the total number of entries in the first
// row of the table is odd, so we need padding
return e[ ((n/4)+1)*4 - 1 ];
}
// it moves vertically across blocks
void kernel_dsymv_4_lib4(int kmax, double *A, int sda, double *x_n, double *y_n, double *z_n, double *x_t, double *y_t, double *z_t, int tri, int alg)
{
if(kmax<=0)
return;
/*printf("\nciao %d\n", kmax); */
const int bs = 4;
__builtin_prefetch( A + bs*0 );
__builtin_prefetch( A + bs*2 );
int k, ka;
ka = kmax; // number from aligned positon
double k_left;
// double *sA, *sy_n, *sx_t;
static double d_mask[4] = {0.5, 1.5, 2.5, 3.5};
__m256d
v_mask,
zeros, temp,
a_00, a_01, a_02, a_03,
x_n_0, x_n_1, x_n_2, x_n_3, y_n_0,
x_t_0, y_t_0, y_t_1, y_t_2, y_t_3;
__m256i
i_mask;
#if 0
__m128d
stemp,
sa_00, sa_01, sa_02, sa_03,
sx_n_0, sx_n_1, sx_n_2, sx_n_3, sy_n_0,
sx_t_0, sy_t_0, sy_t_1, sy_t_2, sy_t_3;
#endif
zeros = _mm256_setzero_pd();
x_n_0 = _mm256_broadcast_sd( &x_n[0] );
x_n_1 = _mm256_broadcast_sd( &x_n[1] );
x_n_2 = _mm256_broadcast_sd( &x_n[2] );
x_n_3 = _mm256_broadcast_sd( &x_n[3] );
if(alg==-1) // TODO xor
{
x_n_0 = _mm256_sub_pd( zeros, x_n_0 );
x_n_1 = _mm256_sub_pd( zeros, x_n_1 );
x_n_2 = _mm256_sub_pd( zeros, x_n_2 );
x_n_3 = _mm256_sub_pd( zeros, x_n_3 );
}
y_t_0 = _mm256_setzero_pd();
y_t_1 = _mm256_setzero_pd();
y_t_2 = _mm256_setzero_pd();
y_t_3 = _mm256_setzero_pd();
#if 0
sx_n_0 = _mm256_castpd256_pd128( x_n_0 );
sx_n_1 = _mm256_castpd256_pd128( x_n_1 );
sx_n_2 = _mm256_castpd256_pd128( x_n_2 );
sx_n_3 = _mm256_castpd256_pd128( x_n_3 );
sy_t_0 = _mm256_castpd256_pd128( y_t_0 );
sy_t_1 = _mm256_castpd256_pd128( y_t_1 );
sy_t_2 = _mm256_castpd256_pd128( y_t_2 );
sy_t_3 = _mm256_castpd256_pd128( y_t_3 );
k = bs*(ka/bs);
sA = A + (ka/bs)*sda*bs;
sy_n = y_n + (ka/bs)*bs;
sx_t = x_t + (ka/bs)*bs;
for(; k<ka; k++)
{
sy_n_0 = _mm_load_sd( &sy_n[0] );
sx_t_0 = _mm_load_sd( &sx_t[0] );
sa_00 = _mm_load_sd( &sA[0+bs*0] );
sa_01 = _mm_load_sd( &sA[0+bs*1] );
sa_02 = _mm_load_sd( &sA[0+bs*2] );
sa_03 = _mm_load_sd( &sA[0+bs*3] );
stemp = _mm_mul_sd( sa_00, sx_n_0 );
sy_n_0 = _mm_add_sd( sy_n_0, stemp );
stemp = _mm_mul_sd( sa_00, sx_t_0 );
sy_t_0 = _mm_add_sd( sy_t_0, stemp );
stemp = _mm_mul_sd( sa_01, sx_n_1 );
sy_n_0 = _mm_add_sd( sy_n_0, stemp );
stemp = _mm_mul_sd( sa_01, sx_t_0 );
sy_t_1 = _mm_add_sd( sy_t_1, stemp );
stemp = _mm_mul_sd( sa_02, sx_n_2 );
sy_n_0 = _mm_add_sd( sy_n_0, stemp );
stemp = _mm_mul_sd( sa_02, sx_t_0 );
sy_t_2 = _mm_add_sd( sy_t_2, stemp );
stemp = _mm_mul_sd( sa_03, sx_n_3 );
sy_n_0 = _mm_add_sd( sy_n_0, stemp );
stemp = _mm_mul_sd( sa_03, sx_t_0 );
sy_t_3 = _mm_add_sd( sy_t_3, stemp );
_mm_store_sd( &sy_n[0], sy_n_0 );
sA += 1;
sy_n += 1;
sx_t += 1;
}
y_t_0 = _mm256_castpd128_pd256( sy_t_0 );
y_t_1 = _mm256_castpd128_pd256( sy_t_1 );
y_t_2 = _mm256_castpd128_pd256( sy_t_2 );
y_t_3 = _mm256_castpd128_pd256( sy_t_3 );
#endif
k=0;
// corner
if(tri==1)
{
__builtin_prefetch( A + sda*bs +bs*0 );
__builtin_prefetch( A + sda*bs +bs*2 );
y_n_0 = _mm256_loadu_pd( &y_n[0] );
x_t_0 = _mm256_loadu_pd( &x_t[0] );
a_00 = _mm256_load_pd( &A[0+bs*0] );
a_01 = _mm256_load_pd( &A[0+bs*1] );
a_02 = _mm256_load_pd( &A[0+bs*2] );
a_03 = _mm256_load_pd( &A[0+bs*3] );
temp = _mm256_mul_pd( a_00, x_n_0 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_00, x_t_0 );
temp = _mm256_blend_pd( zeros, temp, 14 );
y_t_0 = _mm256_add_pd( y_t_0, temp );
temp = _mm256_mul_pd( a_01, x_n_1 );
temp = _mm256_blend_pd( zeros, temp, 14 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_01, x_t_0 );
temp = _mm256_blend_pd( zeros, temp, 12 );
y_t_1 = _mm256_add_pd( y_t_1, temp );
temp = _mm256_mul_pd( a_02, x_n_2 );
temp = _mm256_blend_pd( zeros, temp, 12 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_02, x_t_0 );
temp = _mm256_blend_pd( zeros, temp, 8 );
y_t_2 = _mm256_add_pd( y_t_2, temp );
temp = _mm256_mul_pd( a_03, x_n_3 );
temp = _mm256_blend_pd( zeros, temp, 8 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
_mm256_storeu_pd( &z_n[0], y_n_0 );
A += sda*bs;
y_n += 4;
z_n += 4;
x_t += 4;
k += 4;
}
for(; k<ka-7; k+=2*bs)
{
__builtin_prefetch( A + sda*bs +bs*0 );
__builtin_prefetch( A + sda*bs +bs*2 );
y_n_0 = _mm256_loadu_pd( &y_n[0] );
x_t_0 = _mm256_loadu_pd( &x_t[0] );
a_00 = _mm256_load_pd( &A[0+bs*0] );
a_01 = _mm256_load_pd( &A[0+bs*1] );
a_02 = _mm256_load_pd( &A[0+bs*2] );
a_03 = _mm256_load_pd( &A[0+bs*3] );
temp = _mm256_mul_pd( a_00, x_n_0 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_00, x_t_0 );
y_t_0 = _mm256_add_pd( y_t_0, temp );
temp = _mm256_mul_pd( a_01, x_n_1 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_01, x_t_0 );
y_t_1 = _mm256_add_pd( y_t_1, temp );
temp = _mm256_mul_pd( a_02, x_n_2 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_02, x_t_0 );
y_t_2 = _mm256_add_pd( y_t_2, temp );
temp = _mm256_mul_pd( a_03, x_n_3 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_03, x_t_0 );
y_t_3 = _mm256_add_pd( y_t_3, temp );
_mm256_storeu_pd( &z_n[0], y_n_0 );
A += sda*bs;
y_n += 4;
z_n += 4;
x_t += 4;
__builtin_prefetch( A + sda*bs +bs*0 );
__builtin_prefetch( A + sda*bs +bs*2 );
y_n_0 = _mm256_loadu_pd( &y_n[0] );
x_t_0 = _mm256_loadu_pd( &x_t[0] );
a_00 = _mm256_load_pd( &A[0+bs*0] );
a_01 = _mm256_load_pd( &A[0+bs*1] );
a_02 = _mm256_load_pd( &A[0+bs*2] );
a_03 = _mm256_load_pd( &A[0+bs*3] );
temp = _mm256_mul_pd( a_00, x_n_0 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_00, x_t_0 );
y_t_0 = _mm256_add_pd( y_t_0, temp );
temp = _mm256_mul_pd( a_01, x_n_1 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_01, x_t_0 );
y_t_1 = _mm256_add_pd( y_t_1, temp );
temp = _mm256_mul_pd( a_02, x_n_2 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_02, x_t_0 );
y_t_2 = _mm256_add_pd( y_t_2, temp );
temp = _mm256_mul_pd( a_03, x_n_3 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_03, x_t_0 );
y_t_3 = _mm256_add_pd( y_t_3, temp );
_mm256_storeu_pd( &z_n[0], y_n_0 );
A += sda*bs;
y_n += 4;
z_n += 4;
x_t += 4;
}
for(; k<ka-3; k+=bs)
{
__builtin_prefetch( A + sda*bs +bs*0 );
__builtin_prefetch( A + sda*bs +bs*2 );
y_n_0 = _mm256_loadu_pd( &y_n[0] );
x_t_0 = _mm256_loadu_pd( &x_t[0] );
a_00 = _mm256_load_pd( &A[0+bs*0] );
a_01 = _mm256_load_pd( &A[0+bs*1] );
a_02 = _mm256_load_pd( &A[0+bs*2] );
a_03 = _mm256_load_pd( &A[0+bs*3] );
temp = _mm256_mul_pd( a_00, x_n_0 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_00, x_t_0 );
y_t_0 = _mm256_add_pd( y_t_0, temp );
temp = _mm256_mul_pd( a_01, x_n_1 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_01, x_t_0 );
y_t_1 = _mm256_add_pd( y_t_1, temp );
temp = _mm256_mul_pd( a_02, x_n_2 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_02, x_t_0 );
y_t_2 = _mm256_add_pd( y_t_2, temp );
temp = _mm256_mul_pd( a_03, x_n_3 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_03, x_t_0 );
y_t_3 = _mm256_add_pd( y_t_3, temp );
_mm256_storeu_pd( &z_n[0], y_n_0 );
A += sda*bs;
y_n += 4;
z_n += 4;
x_t += 4;
}
if(k<ka)
{
k_left = ka-k;
v_mask = _mm256_sub_pd( _mm256_loadu_pd( d_mask ), _mm256_broadcast_sd( &k_left ) );
i_mask = _mm256_castpd_si256( v_mask );
// __builtin_prefetch( A + sda*bs +bs*0 );
// __builtin_prefetch( A + sda*bs +bs*2 );
y_n_0 = _mm256_loadu_pd( &y_n[0] );
x_t_0 = _mm256_maskload_pd( &x_t[0], i_mask );
a_00 = _mm256_load_pd( &A[0+bs*0] );
a_01 = _mm256_load_pd( &A[0+bs*1] );
a_02 = _mm256_load_pd( &A[0+bs*2] );
a_03 = _mm256_load_pd( &A[0+bs*3] );
temp = _mm256_mul_pd( a_00, x_n_0 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_00, x_t_0 );
y_t_0 = _mm256_add_pd( y_t_0, temp );
temp = _mm256_mul_pd( a_01, x_n_1 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_01, x_t_0 );
y_t_1 = _mm256_add_pd( y_t_1, temp );
temp = _mm256_mul_pd( a_02, x_n_2 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_02, x_t_0 );
y_t_2 = _mm256_add_pd( y_t_2, temp );
temp = _mm256_mul_pd( a_03, x_n_3 );
y_n_0 = _mm256_add_pd( y_n_0, temp );
temp = _mm256_mul_pd( a_03, x_t_0 );
y_t_3 = _mm256_add_pd( y_t_3, temp );
_mm256_maskstore_pd( &z_n[0], i_mask, y_n_0 );
// A += sda*bs;
// y_n += 4;
// z_n += 4;
// x_t += 4;
}
__m256d
y_0_1_2_3;
y_t_0 = _mm256_hadd_pd( y_t_0, y_t_1 );
y_t_2 = _mm256_hadd_pd( y_t_2, y_t_3 );
y_t_1 = _mm256_permute2f128_pd( y_t_2, y_t_0, 2 );
y_t_0 = _mm256_permute2f128_pd( y_t_2, y_t_0, 19 );
y_t_0 = _mm256_add_pd( y_t_0, y_t_1 );
if(alg==1)
{
y_0_1_2_3 = _mm256_loadu_pd( &y_t[0] );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, y_t_0 );
_mm256_storeu_pd( &z_t[0], y_0_1_2_3 );
}
else // alg==-1
{
y_0_1_2_3 = _mm256_loadu_pd( &y_t[0] );
y_0_1_2_3 = _mm256_sub_pd( y_0_1_2_3, y_t_0 );
_mm256_storeu_pd( &z_t[0], y_0_1_2_3 );
}
}
static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
const unsigned char *_blimit,
const unsigned char *_limit,
const unsigned char *_thresh) {
__m128i mask, hev, flat, flat2;
const __m128i zero = _mm_set1_epi16(0);
const __m128i one = _mm_set1_epi8(1);
__m128i p7, p6, p5;
__m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
__m128i q5, q6, q7;
__m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4, q256_4,
p256_3, q256_3, p256_2, q256_2, p256_1, q256_1, p256_0, q256_0;
const __m128i thresh =
_mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh[0]));
const __m128i limit = _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit[0]));
const __m128i blimit =
_mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit[0]));
p256_4 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 5 * p)));
p256_3 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 4 * p)));
p256_2 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 3 * p)));
p256_1 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 2 * p)));
p256_0 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 1 * p)));
q256_0 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 0 * p)));
q256_1 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 1 * p)));
q256_2 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 2 * p)));
q256_3 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 3 * p)));
q256_4 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 4 * p)));
p4 = _mm256_castsi256_si128(p256_4);
p3 = _mm256_castsi256_si128(p256_3);
p2 = _mm256_castsi256_si128(p256_2);
p1 = _mm256_castsi256_si128(p256_1);
p0 = _mm256_castsi256_si128(p256_0);
q0 = _mm256_castsi256_si128(q256_0);
q1 = _mm256_castsi256_si128(q256_1);
q2 = _mm256_castsi256_si128(q256_2);
q3 = _mm256_castsi256_si128(q256_3);
q4 = _mm256_castsi256_si128(q256_4);
{
const __m128i abs_p1p0 =
_mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1));
const __m128i abs_q1q0 =
_mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1));
const __m128i fe = _mm_set1_epi8(0xfe);
const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
__m128i abs_p0q0 =
_mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0));
__m128i abs_p1q1 =
_mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1));
__m128i work;
flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
hev = _mm_subs_epu8(flat, thresh);
hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
// mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
mask = _mm_max_epu8(flat, mask);
// mask |= (abs(p1 - p0) > limit) * -1;
// mask |= (abs(q1 - q0) > limit) * -1;
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
_mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
mask = _mm_max_epu8(work, mask);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
_mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
mask = _mm_max_epu8(work, mask);
mask = _mm_subs_epu8(mask, limit);
mask = _mm_cmpeq_epi8(mask, zero);
}
// lp filter
{
const __m128i t4 = _mm_set1_epi8(4);
const __m128i t3 = _mm_set1_epi8(3);
const __m128i t80 = _mm_set1_epi8(0x80);
const __m128i te0 = _mm_set1_epi8(0xe0);
const __m128i t1f = _mm_set1_epi8(0x1f);
const __m128i t1 = _mm_set1_epi8(0x1);
const __m128i t7f = _mm_set1_epi8(0x7f);
__m128i ps1 = _mm_xor_si128(p1, t80);
__m128i ps0 = _mm_xor_si128(p0, t80);
__m128i qs0 = _mm_xor_si128(q0, t80);
__m128i qs1 = _mm_xor_si128(q1, t80);
__m128i filt;
__m128i work_a;
__m128i filter1, filter2;
__m128i flat2_p6, flat2_p5, flat2_p4, flat2_p3, flat2_p2, flat2_p1,
flat2_p0, flat2_q0, flat2_q1, flat2_q2, flat2_q3, flat2_q4, flat2_q5,
flat2_q6, flat_p2, flat_p1, flat_p0, flat_q0, flat_q1, flat_q2;
filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
work_a = _mm_subs_epi8(qs0, ps0);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
/* (vpx_filter + 3 * (qs0 - ps0)) & mask */
filt = _mm_and_si128(filt, mask);
filter1 = _mm_adds_epi8(filt, t4);
filter2 = _mm_adds_epi8(filt, t3);
/* Filter1 >> 3 */
work_a = _mm_cmpgt_epi8(zero, filter1);
filter1 = _mm_srli_epi16(filter1, 3);
work_a = _mm_and_si128(work_a, te0);
filter1 = _mm_and_si128(filter1, t1f);
filter1 = _mm_or_si128(filter1, work_a);
qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
/* Filter2 >> 3 */
work_a = _mm_cmpgt_epi8(zero, filter2);
filter2 = _mm_srli_epi16(filter2, 3);
work_a = _mm_and_si128(work_a, te0);
filter2 = _mm_and_si128(filter2, t1f);
filter2 = _mm_or_si128(filter2, work_a);
ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
/* filt >> 1 */
filt = _mm_adds_epi8(filter1, t1);
work_a = _mm_cmpgt_epi8(zero, filt);
filt = _mm_srli_epi16(filt, 1);
work_a = _mm_and_si128(work_a, t80);
filt = _mm_and_si128(filt, t7f);
filt = _mm_or_si128(filt, work_a);
filt = _mm_andnot_si128(hev, filt);
ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
// loopfilter done
{
__m128i work;
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
_mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
flat = _mm_max_epu8(work, flat);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
_mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
flat = _mm_max_epu8(work, flat);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p4, p0), _mm_subs_epu8(p0, p4)),
_mm_or_si128(_mm_subs_epu8(q4, q0), _mm_subs_epu8(q0, q4)));
flat = _mm_subs_epu8(flat, one);
flat = _mm_cmpeq_epi8(flat, zero);
flat = _mm_and_si128(flat, mask);
p256_5 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s - 6 * p)));
q256_5 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s + 5 * p)));
p5 = _mm256_castsi256_si128(p256_5);
q5 = _mm256_castsi256_si128(q256_5);
flat2 = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)),
_mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5)));
flat2 = _mm_max_epu8(work, flat2);
p256_6 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s - 7 * p)));
q256_6 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s + 6 * p)));
p6 = _mm256_castsi256_si128(p256_6);
q6 = _mm256_castsi256_si128(q256_6);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)),
_mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6)));
flat2 = _mm_max_epu8(work, flat2);
p256_7 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s - 8 * p)));
q256_7 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s + 7 * p)));
p7 = _mm256_castsi256_si128(p256_7);
q7 = _mm256_castsi256_si128(q256_7);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)),
_mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7)));
flat2 = _mm_max_epu8(work, flat2);
flat2 = _mm_subs_epu8(flat2, one);
flat2 = _mm_cmpeq_epi8(flat2, zero);
flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// flat and wide flat calculations
{
const __m256i eight = _mm256_set1_epi16(8);
const __m256i four = _mm256_set1_epi16(4);
__m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0,
pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
const __m256i filter =
_mm256_load_si256((__m256i const *)filt_loopfilter_avx2);
p256_7 = _mm256_shuffle_epi8(p256_7, filter);
p256_6 = _mm256_shuffle_epi8(p256_6, filter);
p256_5 = _mm256_shuffle_epi8(p256_5, filter);
p256_4 = _mm256_shuffle_epi8(p256_4, filter);
p256_3 = _mm256_shuffle_epi8(p256_3, filter);
p256_2 = _mm256_shuffle_epi8(p256_2, filter);
p256_1 = _mm256_shuffle_epi8(p256_1, filter);
p256_0 = _mm256_shuffle_epi8(p256_0, filter);
q256_0 = _mm256_shuffle_epi8(q256_0, filter);
q256_1 = _mm256_shuffle_epi8(q256_1, filter);
q256_2 = _mm256_shuffle_epi8(q256_2, filter);
q256_3 = _mm256_shuffle_epi8(q256_3, filter);
q256_4 = _mm256_shuffle_epi8(q256_4, filter);
q256_5 = _mm256_shuffle_epi8(q256_5, filter);
q256_6 = _mm256_shuffle_epi8(q256_6, filter);
q256_7 = _mm256_shuffle_epi8(q256_7, filter);
pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5),
_mm256_add_epi16(p256_4, p256_3));
pixelFilter_q = _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_5),
_mm256_add_epi16(q256_4, q256_3));
pixetFilter_p2p1p0 =
_mm256_add_epi16(p256_0, _mm256_add_epi16(p256_2, p256_1));
pixelFilter_p = _mm256_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
pixetFilter_q2q1q0 =
_mm256_add_epi16(q256_0, _mm256_add_epi16(q256_2, q256_1));
pixelFilter_q = _mm256_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
pixelFilter_p = _mm256_add_epi16(
eight, _mm256_add_epi16(pixelFilter_p, pixelFilter_q));
pixetFilter_p2p1p0 = _mm256_add_epi16(
four, _mm256_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(p256_7, p256_0)), 4);
flat2_p0 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(q256_7, q256_0)), 4);
flat2_q0 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
res_p =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
_mm256_add_epi16(p256_3, p256_0)),
3);
flat_p0 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
_mm256_add_epi16(q256_3, q256_0)),
3);
flat_q0 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(p256_7, p256_7);
sum_q7 = _mm256_add_epi16(q256_7, q256_7);
sum_p3 = _mm256_add_epi16(p256_3, p256_3);
sum_q3 = _mm256_add_epi16(q256_3, q256_3);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_p, p256_6);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_6);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_1)), 4);
flat2_p1 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_1)), 4);
flat2_q1 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_p2p1p0, p256_2);
pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_2);
res_p =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
_mm256_add_epi16(sum_p3, p256_1)),
3);
flat_p1 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0,
_mm256_add_epi16(sum_q3, q256_1)),
3);
flat_q1 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
sum_p3 = _mm256_add_epi16(sum_p3, p256_3);
sum_q3 = _mm256_add_epi16(sum_q3, q256_3);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_5);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_5);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_2)), 4);
flat2_p2 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_2)), 4);
flat2_q2 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_1);
pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_q2q1q0, p256_1);
res_p =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
_mm256_add_epi16(sum_p3, p256_2)),
3);
flat_p2 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0,
_mm256_add_epi16(sum_q3, q256_2)),
3);
flat_q2 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_4);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_4);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_3)), 4);
flat2_p3 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_3)), 4);
flat2_q3 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_3);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_3);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_4)), 4);
flat2_p4 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_4)), 4);
flat2_q4 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_2);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_2);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_5)), 4);
flat2_p5 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_5)), 4);
flat2_q5 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_1);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_1);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_6)), 4);
flat2_p6 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_6)), 4);
flat2_q6 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
}
// wide flat
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
p2 = _mm_andnot_si128(flat, p2);
flat_p2 = _mm_and_si128(flat, flat_p2);
p2 = _mm_or_si128(flat_p2, p2);
p1 = _mm_andnot_si128(flat, ps1);
flat_p1 = _mm_and_si128(flat, flat_p1);
p1 = _mm_or_si128(flat_p1, p1);
p0 = _mm_andnot_si128(flat, ps0);
flat_p0 = _mm_and_si128(flat, flat_p0);
p0 = _mm_or_si128(flat_p0, p0);
q0 = _mm_andnot_si128(flat, qs0);
flat_q0 = _mm_and_si128(flat, flat_q0);
q0 = _mm_or_si128(flat_q0, q0);
q1 = _mm_andnot_si128(flat, qs1);
flat_q1 = _mm_and_si128(flat, flat_q1);
q1 = _mm_or_si128(flat_q1, q1);
q2 = _mm_andnot_si128(flat, q2);
flat_q2 = _mm_and_si128(flat, flat_q2);
q2 = _mm_or_si128(flat_q2, q2);
p6 = _mm_andnot_si128(flat2, p6);
flat2_p6 = _mm_and_si128(flat2, flat2_p6);
p6 = _mm_or_si128(flat2_p6, p6);
_mm_storeu_si128((__m128i *)(s - 7 * p), p6);
p5 = _mm_andnot_si128(flat2, p5);
flat2_p5 = _mm_and_si128(flat2, flat2_p5);
p5 = _mm_or_si128(flat2_p5, p5);
_mm_storeu_si128((__m128i *)(s - 6 * p), p5);
p4 = _mm_andnot_si128(flat2, p4);
flat2_p4 = _mm_and_si128(flat2, flat2_p4);
p4 = _mm_or_si128(flat2_p4, p4);
_mm_storeu_si128((__m128i *)(s - 5 * p), p4);
p3 = _mm_andnot_si128(flat2, p3);
flat2_p3 = _mm_and_si128(flat2, flat2_p3);
p3 = _mm_or_si128(flat2_p3, p3);
_mm_storeu_si128((__m128i *)(s - 4 * p), p3);
p2 = _mm_andnot_si128(flat2, p2);
flat2_p2 = _mm_and_si128(flat2, flat2_p2);
p2 = _mm_or_si128(flat2_p2, p2);
_mm_storeu_si128((__m128i *)(s - 3 * p), p2);
p1 = _mm_andnot_si128(flat2, p1);
flat2_p1 = _mm_and_si128(flat2, flat2_p1);
p1 = _mm_or_si128(flat2_p1, p1);
_mm_storeu_si128((__m128i *)(s - 2 * p), p1);
p0 = _mm_andnot_si128(flat2, p0);
flat2_p0 = _mm_and_si128(flat2, flat2_p0);
p0 = _mm_or_si128(flat2_p0, p0);
_mm_storeu_si128((__m128i *)(s - 1 * p), p0);
q0 = _mm_andnot_si128(flat2, q0);
flat2_q0 = _mm_and_si128(flat2, flat2_q0);
q0 = _mm_or_si128(flat2_q0, q0);
_mm_storeu_si128((__m128i *)(s - 0 * p), q0);
q1 = _mm_andnot_si128(flat2, q1);
flat2_q1 = _mm_and_si128(flat2, flat2_q1);
q1 = _mm_or_si128(flat2_q1, q1);
_mm_storeu_si128((__m128i *)(s + 1 * p), q1);
q2 = _mm_andnot_si128(flat2, q2);
flat2_q2 = _mm_and_si128(flat2, flat2_q2);
q2 = _mm_or_si128(flat2_q2, q2);
_mm_storeu_si128((__m128i *)(s + 2 * p), q2);
q3 = _mm_andnot_si128(flat2, q3);
flat2_q3 = _mm_and_si128(flat2, flat2_q3);
q3 = _mm_or_si128(flat2_q3, q3);
_mm_storeu_si128((__m128i *)(s + 3 * p), q3);
q4 = _mm_andnot_si128(flat2, q4);
flat2_q4 = _mm_and_si128(flat2, flat2_q4);
q4 = _mm_or_si128(flat2_q4, q4);
_mm_storeu_si128((__m128i *)(s + 4 * p), q4);
q5 = _mm_andnot_si128(flat2, q5);
flat2_q5 = _mm_and_si128(flat2, flat2_q5);
q5 = _mm_or_si128(flat2_q5, q5);
_mm_storeu_si128((__m128i *)(s + 5 * p), q5);
q6 = _mm_andnot_si128(flat2, q6);
flat2_q6 = _mm_and_si128(flat2, flat2_q6);
q6 = _mm_or_si128(flat2_q6, q6);
_mm_storeu_si128((__m128i *)(s + 6 * p), q6);
}
}
// Computes and returns the dot product of the n-vectors u and v.
// Uses Intel AVX intrinsics to access the SIMD instruction set.
double DotProductAVX(const double* u, const double* v, int n) {
int max_offset = n - 4;
int offset = 0;
// Accumulate a set of 4 sums in sum, by loading pairs of 4 values from u and
// v, and multiplying them together in parallel.
__m256d sum = _mm256_setzero_pd();
if (offset <= max_offset) {
offset = 4;
// Aligned load is reputedly faster but requires 32 byte aligned input.
if ((reinterpret_cast<const uintptr_t>(u) & 31) == 0 &&
(reinterpret_cast<const uintptr_t>(v) & 31) == 0) {
// Use aligned load.
__m256d floats1 = _mm256_load_pd(u);
__m256d floats2 = _mm256_load_pd(v);
// Multiply.
sum = _mm256_mul_pd(floats1, floats2);
while (offset <= max_offset) {
floats1 = _mm256_load_pd(u + offset);
floats2 = _mm256_load_pd(v + offset);
offset += 4;
__m256d product = _mm256_mul_pd(floats1, floats2);
sum = _mm256_add_pd(sum, product);
}
} else {
// Use unaligned load.
__m256d floats1 = _mm256_loadu_pd(u);
__m256d floats2 = _mm256_loadu_pd(v);
// Multiply.
sum = _mm256_mul_pd(floats1, floats2);
while (offset <= max_offset) {
floats1 = _mm256_loadu_pd(u + offset);
floats2 = _mm256_loadu_pd(v + offset);
offset += 4;
__m256d product = _mm256_mul_pd(floats1, floats2);
sum = _mm256_add_pd(sum, product);
}
}
}
// Add the 4 product sums together horizontally. Not so easy as with sse, as
// there is no add across the upper/lower 128 bit boundary, so permute to
// move the upper 128 bits to lower in another register.
__m256d sum2 = _mm256_permute2f128_pd(sum, sum, 1);
sum = _mm256_hadd_pd(sum, sum2);
sum = _mm256_hadd_pd(sum, sum);
double result;
// _mm256_extract_f64 doesn't exist, but resist the temptation to use an sse
// instruction, as that introduces a 70 cycle delay. All this casting is to
// fool the instrinsics into thinking we are extracting the bottom int64.
auto cast_sum = _mm256_castpd_si256(sum);
*(reinterpret_cast<inT64*>(&result)) =
#if defined(_WIN32) || defined(__i386__)
// This is a very simple workaround that is activated
// for all platforms that do not have _mm256_extract_epi64.
// _mm256_extract_epi64(X, Y) == ((uint64_t*)&X)[Y]
((uint64_t*)&cast_sum)[0]
#else
_mm256_extract_epi64(cast_sum, 0)
#endif
;
while (offset < n) {
result += u[offset] * v[offset];
++offset;
}
return result;
}
