// multiply *p by v and applied to all n
COREARRAY_DLL_DEFAULT void vec_f64_mul(double *p, size_t n, double v)
{
#if defined(COREARRAY_SIMD_AVX)
const __m256d v4 = _mm256_set1_pd(v);
switch ((size_t)p & 0x1F)
{
case 0x08:
if (n > 0) { (*p++) *= v; n--; }
case 0x10:
if (n > 0) { (*p++) *= v; n--; }
case 0x18:
if (n > 0) { (*p++) *= v; n--; }
case 0x00:
for (; n >= 4; n-=4)
{
_mm256_store_pd(p, _mm256_mul_pd(_mm256_load_pd(p), v4));
p += 4;
}
if (n >= 2)
{
_mm_store_pd(p, _mm_mul_pd(_mm_load_pd(p), _mm256_castpd256_pd128(v4)));
p += 2; n -= 2;
}
break;
default:
for (; n >= 4; n-=4)
{
_mm256_storeu_pd(p, _mm256_mul_pd(_mm256_loadu_pd(p), v4));
p += 4;
}
if (n >= 2)
{
_mm_storeu_pd(p, _mm_mul_pd(_mm_loadu_pd(p), _mm256_castpd256_pd128(v4)));
p += 2; n -= 2;
}
}
#elif defined(COREARRAY_SIMD_SSE2)
const __m128d v2 = _mm_set1_pd(v);
switch ((size_t)p & 0x0F)
{
case 0x08:
if (n > 0) { (*p++) *= v; n--; }
case 0x00:
for (; n >= 2; n-=2, p+=2)
_mm_store_pd(p, _mm_mul_pd(_mm_load_pd(p), v2));
break;
default:
for (; n >= 2; n-=2, p+=2)
_mm_storeu_pd(p, _mm_mul_pd(_mm_loadu_pd(p), v2));
}
#endif
for (; n > 0; n--) (*p++) *= v;
}
irreg_poly_area_func_sign(double, _avx) {
if (__builtin_expect(is_null(cords) || cords_len == 0, 0))
return 0;
__m256d
curr,
forw,
coef_0,
coef_1,
end = _mm256_load_pd((const double *)cords),
accum_sum = _mm256_setzero_pd();
double accum_sum_aux;
unsigned long index;
for (index = 0; index < (cords_len - 4); index += 4) {
curr = end; // x0,y0,x1,y1
forw = _mm256_load_pd((const double *)&cords[index + 2]); // x2,y2,x3,y3
end = _mm256_load_pd((const double *)&cords[index + 4]); // x4,y4,x5,y5
coef_0 = _mm256_permute2f128_pd(curr, forw, 0b00110001); // x1, y1, x3, y3
coef_1 = _mm256_permute2f128_pd(forw, end, 0b00100000); // x2, y2, x4, y4
//_mm256_hsub_pd(a, b) == a0 - a1, b0 - b1, a2 - a3, b2 - b3
accum_sum = _mm256_add_pd(
accum_sum,
_mm256_hsub_pd( // x0*y1 - y0*x1, x1*y2 - y1x2, x2*y3 - y2*x3, x3*y4 - y3*x4
_mm256_mul_pd( // x0*y1, y0*x1, x2*y3, y2*x3
_mm256_permute2f128_pd(curr, forw, 0b00100000), // x0, y0, x2, y2
_mm256_shuffle_pd(coef_0, coef_0, 0b0101) // y1, x1, y3, x3
),
_mm256_mul_pd(coef_0, _mm256_shuffle_pd(coef_1, coef_1, 0b0101)) // y2, x2, y4, x4
// ^^^^^^^^^^^^^^^ x1*y2, y1*x2, x3*y4, y3*x4
)
);
}
accum_sum = _mm256_hadd_pd(accum_sum, _mm256_permute2f128_pd(accum_sum, accum_sum, 1)); // a0+a1, a2+a3, a2+a3, a0+a1
accum_sum = _mm256_hadd_pd(accum_sum, accum_sum); // a0+a1+a2+a3, ...
for (accum_sum_aux = _mm_cvtsd_f64(_mm256_castpd256_pd128(accum_sum)); index < (cords_len - 1); index++)
accum_sum_aux += _calc_diff_of_adj_prods(cords, index);
return accum_sum_aux;
// return scalar_half(scalar_abs(accum_sum_aux));
}
/*!
* \brief Perform an horizontal sum of the given vector.
* \param in The input vector type
* \return the horizontal sum of the vector
*/
ETL_STATIC_INLINE(double) hadd(avx_simd_double in) {
const __m256d t1 = _mm256_hadd_pd(in.value, _mm256_permute2f128_pd(in.value, in.value, 1));
const __m256d t2 = _mm256_hadd_pd(t1, t1);
return _mm_cvtsd_f64(_mm256_castpd256_pd128(t2));
}
void AVX2FMA3DNoise(Vector3d& result, const Vector3d& EPoint)
{
#if CHECK_FUNCTIONAL
Vector3d param(EPoint);
#endif
AVX2TABLETYPE *mp;
// TODO FIXME - global statistics reference
// Stats[Calls_To_DNoise]++;
const __m256d ONE_PD = _mm256_set1_pd(1.0);
const __m128i short_si128 = _mm_set1_epi32(0xffff);
const __m256d xyzn = _mm256_setr_pd(EPoint[X], EPoint[Y], EPoint[Z], 0);
const __m256d epsy = _mm256_set1_pd(1.0 - EPSILON);
const __m256d xyzn_e = _mm256_sub_pd(xyzn, epsy);
const __m128i tmp_xyzn = _mm256_cvttpd_epi32(_mm256_blendv_pd(xyzn, xyzn_e, xyzn));
const __m128i noise_min_xyzn = _mm_setr_epi32(NOISE_MINX, NOISE_MINY, NOISE_MINZ, 0);
const __m256d xyz_ixyzn = _mm256_sub_pd(xyzn, _mm256_cvtepi32_pd(tmp_xyzn));
const __m256d xyz_jxyzn = _mm256_sub_pd(xyz_ixyzn, ONE_PD);
const __m128i i_xyzn = _mm_and_si128(_mm_sub_epi32(tmp_xyzn, noise_min_xyzn),
_mm_set1_epi32(0xfff));
const __m256d s_xyzn = _mm256_mul_pd(xyz_ixyzn,
_mm256_mul_pd(xyz_ixyzn,
_mm256_sub_pd(_mm256_set1_pd(3.0),
_mm256_add_pd(xyz_ixyzn, xyz_ixyzn))));
const __m256d t_xyzn = _mm256_sub_pd(ONE_PD, s_xyzn);
const __m256d txtysxsy = _mm256_permute2f128_pd(t_xyzn, s_xyzn, 0x20);
const __m256d txsxtxsx = PERMUTE4x64(txtysxsy, _MM_SHUFFLE(2, 0, 2, 0));
const __m256d tytysysy = PERMUTE4x64(txtysxsy, _MM_SHUFFLE(3, 3, 1, 1));
const __m256d txtysxtytxsysxsy = _mm256_mul_pd(txsxtxsx, tytysysy);
const __m256d incrsump_s1 = _mm256_mul_pd(txtysxtytxsysxsy, PERMUTE4x64(t_xyzn, _MM_SHUFFLE(2, 2, 2, 2)));
const __m256d incrsump_s2 = _mm256_mul_pd(txtysxtytxsysxsy, PERMUTE4x64(s_xyzn, _MM_SHUFFLE(2, 2, 2, 2)));
int ints[4];
_mm_storeu_si128((__m128i*)(ints), i_xyzn);
const int ixiy_hash = Hash2d(ints[0], ints[1]);
const int jxiy_hash = Hash2d(ints[0] + 1, ints[1]);
const int ixjy_hash = Hash2d(ints[0], ints[1] + 1);
const int jxjy_hash = Hash2d(ints[0] + 1, ints[1] + 1);
const int iz = ints[2];
const __m256d iii = _mm256_blend_pd(PERMUTE4x64(xyz_ixyzn, _MM_SHUFFLE(2, 1, 0, 0)), _mm256_set_pd(0, 0, 0, 0.5), 0x1);
const __m256d jjj = _mm256_blend_pd(PERMUTE4x64(xyz_jxyzn, _MM_SHUFFLE(2, 1, 0, 0)), _mm256_set_pd(0, 0, 0, 0.5), 0x1);
__m256d ss;
__m256d blend;
__m256d x = _mm256_setzero_pd(), y = _mm256_setzero_pd(), z = _mm256_setzero_pd();
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixiy_hash, iz)];
ss = PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(0, 0, 0, 0));
// blend = _mm256_blend_pd(iii, jjj, 0);
INCSUMAVX_VECTOR(mp, ss, iii);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxiy_hash, iz)];
ss = PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(1, 1, 1, 1));
blend = _mm256_blend_pd(iii, jjj, 2);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxjy_hash, iz)];
ss = PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(3, 3, 3, 3));
blend = _mm256_blend_pd(iii, jjj, 6);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixjy_hash, iz)];
ss = PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(2, 2, 2, 2));
blend = _mm256_blend_pd(iii, jjj, 4);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixjy_hash, iz + 1)];
ss = PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(2, 2, 2, 2));
blend = _mm256_blend_pd(iii, jjj, 12);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxjy_hash, iz + 1)];
ss = PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(3, 3, 3, 3));
// blend = _mm256_blend_pd(iii, jjj, 14);
INCSUMAVX_VECTOR(mp, ss, jjj);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxiy_hash, iz + 1)];
ss = PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(1, 1, 1, 1));
blend = _mm256_blend_pd(iii, jjj, 10);
INCSUMAVX_VECTOR(mp, ss, blend);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixiy_hash, iz + 1)];
ss = PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(0, 0, 0, 0));
blend = _mm256_blend_pd(iii, jjj, 8);
INCSUMAVX_VECTOR(mp, ss, blend);
__m256d xy = _mm256_hadd_pd(x,y);
__m128d xy_up = _mm256_extractf128_pd(xy,1);
xy_up = _mm_add_pd(_mm256_castpd256_pd128(xy),xy_up);
_mm_storeu_pd(&result[X],xy_up);
__m128d z_up = _mm256_extractf128_pd(z,1);
z_up = _mm_add_pd(_mm256_castpd256_pd128(z),z_up);
z_up = _mm_hadd_pd(z_up,z_up);
result[Z] = _mm_cvtsd_f64(z_up);
#if CHECK_FUNCTIONAL
{
Vector3d portable_res;
PortableDNoise(portable_res , param);
if (fabs(portable_res[X] - result[X]) >= EPSILON)
{
throw POV_EXCEPTION_STRING("DNoise X error");
}
if (fabs(portable_res[Y] - result[Y]) >= EPSILON)
{
throw POV_EXCEPTION_STRING("DNoise Y error");
}
if (fabs(portable_res[Z] - result[Z]) >= EPSILON)
{
throw POV_EXCEPTION_STRING("DNoise Z error");
}
}
#endif
_mm256_zeroupper();
return;
}
DBL AVX2FMA3Noise(const Vector3d& EPoint, int noise_generator)
{
AVX2TABLETYPE *mp;
DBL sum = 0.0;
// TODO FIXME - global statistics reference
// Stats[Calls_To_Noise]++;
if (noise_generator == kNoiseGen_Perlin)
{
// The 1.59 and 0.985 are to correct for some biasing problems with
// the random # generator used to create the noise tables. Final
// range of values is about 5.0e-4 below 0.0 and above 1.0. Mean
// value is 0.49 (ideally it would be 0.5).
sum = 0.5 * (1.59 * SolidNoise(EPoint) + 0.985);
// Clamp final value to 0-1 range
if (sum < 0.0) sum = 0.0;
if (sum > 1.0) sum = 1.0;
return sum;
}
const __m256d ONE_PD = _mm256_set1_pd(1);
const __m128i short_si128 = _mm_set1_epi32(0xffff);
const __m256d xyzn = _mm256_setr_pd(EPoint[X], EPoint[Y], EPoint[Z], 0);
const __m256d epsy = _mm256_set1_pd(1.0 - EPSILON);
const __m256d xyzn_e = _mm256_sub_pd(xyzn, epsy);
const __m128i tmp_xyzn = _mm256_cvttpd_epi32(_mm256_blendv_pd(xyzn, xyzn_e, xyzn));
const __m128i noise_min_xyzn = _mm_setr_epi32(NOISE_MINX, NOISE_MINY, NOISE_MINZ, 0);
const __m256d xyz_ixyzn = _mm256_sub_pd(xyzn, _mm256_cvtepi32_pd(tmp_xyzn));
const __m256d xyz_jxyzn = _mm256_sub_pd(xyz_ixyzn, ONE_PD);
const __m128i i_xyzn = _mm_and_si128(_mm_sub_epi32(tmp_xyzn, noise_min_xyzn),
_mm_set1_epi32(0xfff));
const __m256d s_xyzn = _mm256_mul_pd(xyz_ixyzn,
_mm256_mul_pd(xyz_ixyzn,
_mm256_sub_pd(_mm256_set1_pd(3.0),
_mm256_add_pd(xyz_ixyzn, xyz_ixyzn))));
const __m256d t_xyzn = _mm256_sub_pd(ONE_PD, s_xyzn);
const __m256d txtysxsy = _mm256_permute2f128_pd(t_xyzn, s_xyzn, 0x20);
const __m256d txsxtxsx = PERMUTE4x64(txtysxsy, _MM_SHUFFLE(2, 0, 2, 0));
const __m256d tytysysy = PERMUTE4x64(txtysxsy, _MM_SHUFFLE(3, 3, 1, 1));
const __m256d txtysxtytxsysxsy = _mm256_mul_pd(txsxtxsx, tytysysy);
const __m256d incrsump_s1 = _mm256_mul_pd(txtysxtytxsysxsy, PERMUTE4x64(t_xyzn, _MM_SHUFFLE(2, 2, 2, 2)));
const __m256d incrsump_s2 = _mm256_mul_pd(txtysxtytxsysxsy, PERMUTE4x64(s_xyzn, _MM_SHUFFLE(2, 2, 2, 2)));
int ints[4];
_mm_storeu_si128((__m128i*)(ints), i_xyzn);
const int ixiy_hash = Hash2d(ints[0], ints[1]);
const int jxiy_hash = Hash2d(ints[0] + 1, ints[1]);
const int ixjy_hash = Hash2d(ints[0], ints[1] + 1);
const int jxjy_hash = Hash2d(ints[0] + 1, ints[1] + 1);
const int iz = ints[2];
const __m256d iii = _mm256_blend_pd(PERMUTE4x64(xyz_ixyzn, _MM_SHUFFLE(2, 1, 0, 0)), _mm256_set_pd(0, 0, 0, 0.5), 0x1);
const __m256d jjj = _mm256_blend_pd(PERMUTE4x64(xyz_jxyzn, _MM_SHUFFLE(2, 1, 0, 0)), _mm256_set_pd(0, 0, 0, 0.5), 0x1);
__m256d sumr = _mm256_setzero_pd();
__m256d sumr1 = _mm256_setzero_pd();
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixiy_hash, iz)];
INCSUMAVX_NOBLEND(sumr, mp, PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(0, 0, 0, 0)), iii);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxiy_hash, iz)];
INCSUMAVX(sumr1, mp, PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(1, 1, 1, 1)), iii, jjj, 2);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixjy_hash, iz)];
INCSUMAVX(sumr, mp, PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(2, 2, 2, 2)), iii, jjj, 4);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxjy_hash, iz)];
INCSUMAVX(sumr1, mp, PERMUTE4x64(incrsump_s1, _MM_SHUFFLE(3, 3, 3, 3)), iii, jjj, 6);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixiy_hash, iz + 1)];
INCSUMAVX(sumr, mp, PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(0, 0, 0, 0)), iii, jjj, 8);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxiy_hash, iz + 1)];
INCSUMAVX(sumr1, mp, PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(1, 1, 1, 1)), iii, jjj, 10);
mp = &AVX2RTable[Hash1dRTableIndexAVX(ixjy_hash, iz + 1)];
INCSUMAVX(sumr, mp, PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(2, 2, 2, 2)), iii, jjj, 12);
mp = &AVX2RTable[Hash1dRTableIndexAVX(jxjy_hash, iz + 1)];
INCSUMAVX_NOBLEND(sumr1, mp, PERMUTE4x64(incrsump_s2, _MM_SHUFFLE(3, 3, 3, 3)), jjj);
{
sumr = _mm256_add_pd(sumr, sumr1);
__m128d sumr_up = _mm256_extractf128_pd(sumr,1);
sumr_up = _mm_add_pd(_mm256_castpd256_pd128(sumr),sumr_up);
sumr_up = _mm_hadd_pd(sumr_up,sumr_up);
sum = _mm_cvtsd_f64(sumr_up);
}
if (noise_generator == kNoiseGen_RangeCorrected)
{
/* details of range here:
Min, max: -1.05242, 0.988997
Mean: -0.0191481, Median: -0.535493, Std Dev: 0.256828
We want to change it to as close to [0,1] as possible.
*/
sum += 1.05242;
sum *= 0.48985582;
/*sum *= 0.5;
sum += 0.5;*/
if (sum < 0.0)
sum = 0.0;
if (sum > 1.0)
sum = 1.0;
}
else
{
sum = sum + 0.5; /* range at this point -0.5 - 0.5... */
if (sum < 0.0)
sum = 0.0;
if (sum > 1.0)
sum = 1.0;
}
#if CHECK_FUNCTIONAL
{
DBL orig_sum = PortableNoise(EPoint, noise_generator);
if (fabs(orig_sum - sum) >= EPSILON)
{
throw POV_EXCEPTION_STRING("Noise error");
}
}
#endif
_mm256_zeroupper();
return (sum);
}
inline void rotate_left_wm2(F64vec4 *v0, const F64vec4 v1)
{
*v0 = _mm256_castpd128_pd256(_mm256_extractf128_pd(*v0, 1));
*v0 = _mm256_insertf128_pd(*v0, _mm256_castpd256_pd128(v1), 1);
}
// it moves vertically across blocks
void kernel_dtrmv_u_t_1_lib4(int kmax, double *A, int sda, double *x, double *y, int alg)
{
/* if(kmax<=0) */
/* return;*/
const int lda = 4;
double *tA, *tx;
int k;
__m256d
tmp0,
a_00_10_20_30,
x_0_1_2_3,
y_00;
y_00 = _mm256_setzero_pd();
k=0;
for(; k<kmax-3; k+=4)
{
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, tmp0 );
A += 4 + (sda-1)*lda;
x += 4;
}
__m128d
tm0,
a_00_10, a_01_11,
x_0_1,
y_0, y_1, y_0_1;
tm0 = _mm256_extractf128_pd( y_00, 0x1 );
y_0 = _mm256_castpd256_pd128( y_00 );
y_0 = _mm_add_pd( y_0, tm0 );
if(k<kmax-1)
{
x_0_1 = _mm_loadu_pd( &x[0] );
a_00_10 = _mm_load_pd( &A[0+lda*0] );
tm0 = _mm_mul_pd( a_00_10, x_0_1 );
y_0 = _mm_add_pd( y_0, tm0 );
A += 2;
x += 2;
}
x_0_1 = _mm_load_sd( &x[0] );
a_00_10 = _mm_load_sd( &A[0+lda*0] );
tm0 = _mm_mul_sd( a_00_10, x_0_1 );
y_0 = _mm_add_sd( y_0, tm0 );
y_0 = _mm_hadd_pd( y_0, y_0 );
if(alg==0)
{
_mm_store_sd(&y[0], y_0);
}
else if(alg==1)
{
y_0_1 = _mm_load_sd( &y[0] );
y_0_1 = _mm_add_sd( y_0_1, y_0 );
_mm_store_sd(&y[0], y_0_1);
}
else // alg==-1
{
y_0_1 = _mm_load_sd( &y[0] );
y_0_1 = _mm_sub_sd( y_0_1, y_0 );
_mm_store_sd(&y[0], y_0_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 );
}
}
void kernel_dgemv_t_1_lib4(int kmax, double *A, int sda, double *x, double *y, int alg)
{
if(kmax<=0)
return;
const int lda = 4;
double *tA, *tx;
int k;
int ka = kmax; // number from aligned positon
__m256d
aaxx_temp,
a_00_10_20_30,
x_0_1_2_3,
y_00;
__m128d
ax_temp,
a_00_10,
x_0_1,
y_0, y_1, y_0_1;
y_00 = _mm256_setzero_pd();
y_0 = _mm256_castpd256_pd128(y_00);
k = lda*(ka/lda);
tA = A + (ka/lda)*sda*lda;
tx = x + (ka/lda)*lda;
for(; k<ka; k++)
{
x_0_1 = _mm_load_sd( &tx[0] );
a_00_10 = _mm_load_sd( &tA[0+lda*0] );
/* y_0 += a_00_10 * x_0_1;*/
ax_temp = _mm_mul_sd( a_00_10, x_0_1 );
y_0 = _mm_add_sd (y_0, ax_temp );
tA += 1;
tx += 1;
}
y_00 = _mm256_castpd128_pd256(y_0);
k=0;
for(; k<ka-3; k+=4)
{
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
/* y_00 += a_00_10_20_30 * x_0_1_2_3;*/
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, aaxx_temp );
A += 4 + (sda-1)*lda;
x += 4;
}
y_00 = _mm256_hadd_pd(y_00, y_00);
y_1 = _mm256_extractf128_pd(y_00, 1);
y_0 = _mm256_castpd256_pd128(y_00);
/* y_0 += y_1;*/
y_0 = _mm_add_sd( y_0, y_1 );
if(alg==0)
{
_mm_store_sd(&y[0], y_0);
}
else if(alg==1)
{
y_0_1 = _mm_load_sd( &y[0] );
/* y_0_1 += y_0;*/
y_0_1 = _mm_add_sd( y_0_1, y_0 );
_mm_store_sd(&y[0], y_0_1);
}
else // alg==-1
{
y_0_1 = _mm_load_sd( &y[0] );
/* y_0_1 -= y_0;*/
y_0_1 = _mm_sub_sd( y_0_1, y_0 );
_mm_store_sd(&y[0], y_0_1);
}
}
void kernel_dgemv_t_8_lib4(int kmax, double *A, int sda, double *x, double *y, int alg)
{
if(kmax<=0)
return;
const int lda = 4;
__builtin_prefetch( A + 0*lda );
__builtin_prefetch( A + 2*lda );
__builtin_prefetch( A + 4*lda );
__builtin_prefetch( A + 6*lda );
double *tA, *tx;
int k;
int ka = kmax; // number from aligned positon
__m256d
aaxx_temp,
a_00_10_20_30, a_01_11_21_31, a_02_12_22_32, a_03_13_23_33,
x_0_1_2_3,
y_00, y_11, y_22, y_33, y_44, y_55, y_66, y_77;
__m128d
ax_temp,
a_00_10, a_01_11, a_02_12, a_03_13,
x_0_1,
y_0, y_1, y_2, y_3, y_4, y_5, y_6, y_7;
y_00 = _mm256_setzero_pd();
y_11 = _mm256_setzero_pd();
y_22 = _mm256_setzero_pd();
y_33 = _mm256_setzero_pd();
y_44 = _mm256_setzero_pd();
y_55 = _mm256_setzero_pd();
y_66 = _mm256_setzero_pd();
y_77 = _mm256_setzero_pd();
y_0 = _mm256_castpd256_pd128(y_00);
y_1 = _mm256_castpd256_pd128(y_11);
y_2 = _mm256_castpd256_pd128(y_22);
y_3 = _mm256_castpd256_pd128(y_33);
y_4 = _mm256_castpd256_pd128(y_44);
y_5 = _mm256_castpd256_pd128(y_55);
y_6 = _mm256_castpd256_pd128(y_66);
y_7 = _mm256_castpd256_pd128(y_77);
k = lda*(ka/lda);
tA = A + (ka/lda)*sda*lda;
tx = x + (ka/lda)*lda;
if(ka-k>0) // it can be only ka-k = {1, 2, 3}
{
if((ka-k)>=2)
{
x_0_1 = _mm_load_pd( &tx[0] );
a_00_10 = _mm_load_pd( &tA[0+lda*0] );
a_01_11 = _mm_load_pd( &tA[0+lda*1] );
a_02_12 = _mm_load_pd( &tA[0+lda*2] );
a_03_13 = _mm_load_pd( &tA[0+lda*3] );
ax_temp = _mm_mul_pd( a_00_10, x_0_1 );
y_0 = _mm_add_pd (y_0, ax_temp );
ax_temp = _mm_mul_pd( a_01_11, x_0_1 );
y_1 = _mm_add_pd (y_1, ax_temp );
ax_temp = _mm_mul_pd( a_02_12, x_0_1 );
y_2 = _mm_add_pd (y_2, ax_temp );
ax_temp = _mm_mul_pd( a_03_13, x_0_1 );
y_3 = _mm_add_pd (y_3, ax_temp );
a_00_10 = _mm_load_pd( &tA[0+lda*4] );
a_01_11 = _mm_load_pd( &tA[0+lda*5] );
a_02_12 = _mm_load_pd( &tA[0+lda*6] );
a_03_13 = _mm_load_pd( &tA[0+lda*7] );
ax_temp = _mm_mul_pd( a_00_10, x_0_1 );
y_4 = _mm_add_pd (y_4, ax_temp );
ax_temp = _mm_mul_pd( a_01_11, x_0_1 );
y_5 = _mm_add_pd (y_5, ax_temp );
ax_temp = _mm_mul_pd( a_02_12, x_0_1 );
y_6 = _mm_add_pd (y_6, ax_temp );
ax_temp = _mm_mul_pd( a_03_13, x_0_1 );
y_7 = _mm_add_pd (y_7, ax_temp );
tA += 2;
tx += 2;
k+=2;
}
if((ka-k)==1)
{
x_0_1 = _mm_load_sd( &tx[0] );
a_00_10 = _mm_load_sd( &tA[0+lda*0] );
a_01_11 = _mm_load_sd( &tA[0+lda*1] );
a_02_12 = _mm_load_sd( &tA[0+lda*2] );
a_03_13 = _mm_load_sd( &tA[0+lda*3] );
ax_temp = _mm_mul_sd( a_00_10, x_0_1 );
y_0 = _mm_add_sd (y_0, ax_temp );
ax_temp = _mm_mul_sd( a_01_11, x_0_1 );
y_1 = _mm_add_sd (y_1, ax_temp );
ax_temp = _mm_mul_sd( a_02_12, x_0_1 );
y_2 = _mm_add_sd (y_2, ax_temp );
ax_temp = _mm_mul_sd( a_03_13, x_0_1 );
y_3 = _mm_add_sd (y_3, ax_temp );
a_00_10 = _mm_load_sd( &tA[0+lda*4] );
a_01_11 = _mm_load_sd( &tA[0+lda*5] );
a_02_12 = _mm_load_sd( &tA[0+lda*6] );
a_03_13 = _mm_load_sd( &tA[0+lda*7] );
ax_temp = _mm_mul_sd( a_00_10, x_0_1 );
y_4 = _mm_add_sd (y_4, ax_temp );
ax_temp = _mm_mul_sd( a_01_11, x_0_1 );
y_5 = _mm_add_sd (y_5, ax_temp );
ax_temp = _mm_mul_sd( a_02_12, x_0_1 );
y_6 = _mm_add_sd (y_6, ax_temp );
ax_temp = _mm_mul_sd( a_03_13, x_0_1 );
y_7 = _mm_add_sd (y_7, ax_temp );
tA += 1;
tx += 1;
k++;
}
}
y_00 = _mm256_castpd128_pd256(y_0);
y_11 = _mm256_castpd128_pd256(y_1);
y_22 = _mm256_castpd128_pd256(y_2);
y_33 = _mm256_castpd128_pd256(y_3);
y_44 = _mm256_castpd128_pd256(y_4);
y_55 = _mm256_castpd128_pd256(y_5);
y_66 = _mm256_castpd128_pd256(y_6);
y_77 = _mm256_castpd128_pd256(y_7);
k=0;
for(; k<ka-7; k+=8)
{
__builtin_prefetch( A + sda*lda + 0*lda );
__builtin_prefetch( A + sda*lda + 2*lda );
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_11 = _mm256_add_pd( y_11, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_33 = _mm256_add_pd( y_33, aaxx_temp );
__builtin_prefetch( A + sda*lda + 4*lda );
__builtin_prefetch( A + sda*lda + 6*lda );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*4] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*5] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*6] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*7] );
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_44 = _mm256_add_pd( y_44, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_55 = _mm256_add_pd( y_55, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
y_66 = _mm256_add_pd( y_66, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_77 = _mm256_add_pd( y_77, aaxx_temp );
A += 4 + (sda-1)*lda;
x += 4;
__builtin_prefetch( A + sda*lda + 0*lda );
__builtin_prefetch( A + sda*lda + 2*lda );
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_11 = _mm256_add_pd( y_11, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_33 = _mm256_add_pd( y_33, aaxx_temp );
__builtin_prefetch( A + sda*lda + 4*lda );
__builtin_prefetch( A + sda*lda + 6*lda );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*4] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*5] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*6] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*7] );
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_44 = _mm256_add_pd( y_44, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_55 = _mm256_add_pd( y_55, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
y_66 = _mm256_add_pd( y_66, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_77 = _mm256_add_pd( y_77, aaxx_temp );
A += 4 + (sda-1)*lda;
x += 4;
}
for(; k<ka-3; k+=4)
{
__builtin_prefetch( A + sda*lda + 0*lda );
__builtin_prefetch( A + sda*lda + 2*lda );
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_11 = _mm256_add_pd( y_11, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_33 = _mm256_add_pd( y_33, aaxx_temp );
__builtin_prefetch( A + sda*lda + 4*lda );
__builtin_prefetch( A + sda*lda + 6*lda );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*4] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*5] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*6] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*7] );
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_44 = _mm256_add_pd( y_44, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_55 = _mm256_add_pd( y_55, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
y_66 = _mm256_add_pd( y_66, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_77 = _mm256_add_pd( y_77, aaxx_temp );
A += 4 + (sda-1)*lda;
x += 4;
}
__m256d
y_0_1_2_3, y_4_5_6_7;
y_00 = _mm256_hadd_pd(y_00, y_11);
y_22 = _mm256_hadd_pd(y_22, y_33);
y_44 = _mm256_hadd_pd(y_44, y_55);
y_66 = _mm256_hadd_pd(y_66, y_77);
y_11 = _mm256_permute2f128_pd(y_22, y_00, 2 );
y_00 = _mm256_permute2f128_pd(y_22, y_00, 19);
y_55 = _mm256_permute2f128_pd(y_66, y_44, 2 );
y_44 = _mm256_permute2f128_pd(y_66, y_44, 19);
y_00 = _mm256_add_pd( y_00, y_11 );
y_44 = _mm256_add_pd( y_44, y_55 );
if(alg==0)
{
_mm256_storeu_pd(&y[0], y_00);
_mm256_storeu_pd(&y[4], y_44);
}
else if(alg==1)
{
y_0_1_2_3 = _mm256_loadu_pd( &y[0] );
y_4_5_6_7 = _mm256_loadu_pd( &y[4] );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, y_00 );
y_4_5_6_7 = _mm256_add_pd( y_4_5_6_7, y_44 );
_mm256_storeu_pd(&y[0], y_0_1_2_3);
_mm256_storeu_pd(&y[4], y_4_5_6_7);
}
else // alg==-1
{
y_0_1_2_3 = _mm256_loadu_pd( &y[0] );
y_4_5_6_7 = _mm256_loadu_pd( &y[4] );
y_0_1_2_3 = _mm256_sub_pd( y_0_1_2_3, y_00 );
y_4_5_6_7 = _mm256_sub_pd( y_4_5_6_7, y_44 );
_mm256_storeu_pd(&y[0], y_0_1_2_3);
_mm256_storeu_pd(&y[4], y_4_5_6_7);
}
}
void kernel_dgemv_t_4_lib4(int kmax, double *A, int sda, double *x, double *y, int alg)
{
if(kmax<=0)
return;
const int lda = 4;
__builtin_prefetch( A + 0*lda );
__builtin_prefetch( A + 2*lda );
double *tA, *tx;
int k;
int ka = kmax; // number from aligned positon
__m256d
aaxx_temp,
a_00_10_20_30, a_01_11_21_31, a_02_12_22_32, a_03_13_23_33,
x_0_1_2_3,
y_00, y_11, y_22, y_33;
__m128d
ax_temp,
a_00_10, a_01_11, a_02_12, a_03_13,
x_0_1,
y_0, y_1, y_2, y_3;
y_00 = _mm256_setzero_pd();
y_11 = _mm256_setzero_pd();
y_22 = _mm256_setzero_pd();
y_33 = _mm256_setzero_pd();
y_0 = _mm256_castpd256_pd128(y_00);
y_1 = _mm256_castpd256_pd128(y_11);
y_2 = _mm256_castpd256_pd128(y_22);
y_3 = _mm256_castpd256_pd128(y_33);
k = lda*(ka/lda);
tA = A + (ka/lda)*sda*lda;
tx = x + (ka/lda)*lda;
for(; k<ka; k++)
{
x_0_1 = _mm_load_sd( &tx[0] );
a_00_10 = _mm_load_sd( &tA[0+lda*0] );
a_01_11 = _mm_load_sd( &tA[0+lda*1] );
a_02_12 = _mm_load_sd( &tA[0+lda*2] );
a_03_13 = _mm_load_sd( &tA[0+lda*3] );
ax_temp = _mm_mul_sd( a_00_10, x_0_1 );
y_0 = _mm_add_sd (y_0, ax_temp );
ax_temp = _mm_mul_sd( a_01_11, x_0_1 );
y_1 = _mm_add_sd (y_1, ax_temp );
ax_temp = _mm_mul_sd( a_02_12, x_0_1 );
y_2 = _mm_add_sd (y_2, ax_temp );
ax_temp = _mm_mul_sd( a_03_13, x_0_1 );
y_3 = _mm_add_sd (y_3, ax_temp );
tA += 1;
tx += 1;
}
y_00 = _mm256_castpd128_pd256(y_0);
y_11 = _mm256_castpd128_pd256(y_1);
y_22 = _mm256_castpd128_pd256(y_2);
y_33 = _mm256_castpd128_pd256(y_3);
k=0;
for(; k<ka-7; k+=8)
{
__builtin_prefetch( A + sda*lda + 0*lda );
__builtin_prefetch( A + sda*lda + 2*lda );
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_11 = _mm256_add_pd( y_11, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_33 = _mm256_add_pd( y_33, aaxx_temp );
A += 4 + (sda-1)*lda;
x += 4;
__builtin_prefetch( A + sda*lda + 0*lda );
__builtin_prefetch( A + sda*lda + 2*lda );
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_11 = _mm256_add_pd( y_11, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_33 = _mm256_add_pd( y_33, aaxx_temp );
A += 4 + (sda-1)*lda;
x += 4;
}
for(; k<ka-3; k+=4)
{
__builtin_prefetch( A + sda*lda + 0*lda );
__builtin_prefetch( A + sda*lda + 2*lda );
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
aaxx_temp = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_11 = _mm256_add_pd( y_11, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, aaxx_temp );
aaxx_temp = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_33 = _mm256_add_pd( y_33, aaxx_temp );
A += 4 + (sda-1)*lda;
x += 4;
}
__m256d
y_0_1_2_3;
y_00 = _mm256_hadd_pd(y_00, y_11);
y_22 = _mm256_hadd_pd(y_22, y_33);
y_11 = _mm256_permute2f128_pd(y_22, y_00, 2 );
y_00 = _mm256_permute2f128_pd(y_22, y_00, 19);
y_00 = _mm256_add_pd( y_00, y_11 );
if(alg==0)
{
_mm256_storeu_pd(&y[0], y_00);
}
else if(alg==1)
{
y_0_1_2_3 = _mm256_loadu_pd( &y[0] );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, y_00 );
_mm256_storeu_pd(&y[0], y_0_1_2_3);
}
else // alg==-1
{
y_0_1_2_3 = _mm256_loadu_pd( &y[0] );
y_0_1_2_3 = _mm256_sub_pd( y_0_1_2_3, y_00 );
_mm256_storeu_pd(&y[0], y_0_1_2_3);
}
}
