inline void rotate_left_wm1(F64vec4 *v0, const F64vec4 v1)
{
// {1.0, 2.0, 3.0, 4.0};
// {5.0, 6.0, 7.0, 8.0};
const __m128d hiv0 = _mm256_extractf128_pd(*v0, 1); // {3.0, 4.0}
const __m128d phiv0 = _mm_permute_pd(hiv0, 0x1); // {4.0, 3.0}
const __m256d shufv1 = _mm256_permute_pd(v1, 0x1); // {6.0, 5.0, 8.0, 7.0};
const __m128d shufv1_lo = _mm256_extractf128_pd(shufv1, 0); // {6.0, 5.0}
const __m128d shufv1_hi = _mm256_extractf128_pd(shufv1, 1); // {8.0, 7.0}
const __m128d v1_blend = _mm_blend_pd(shufv1_lo, shufv1_hi, 0x2); // blend {6.0, 7.0};
const __m256d inserted = _mm256_insertf128_pd(shufv1, v1_blend, 1); // insert {6.0, 5.0, 6.0, 7.0};
const __m256d blended = _mm256_blend_pd(_mm256_castpd128_pd256(phiv0), inserted, 0xE);
*v0 = blended;
}
// it moves horizontally inside a block (A upper triangular)
void kernel_dtrmv_u_n_4_lib4(int kmax, double *A, double *x, double *y, int alg)
{
if(kmax<=0)
return;
const int lda = 4;
int k;
__m128d
tmp0,
z_0, y_0_1, a_00_10;
__m256d
zeros,
ax_temp,
a_00_10_20_30, a_01_11_21_31, a_02_12_22_32, a_03_13_23_33,
x_0, x_1, x_2, x_3,
y_0_1_2_3, y_0_1_2_3_b, y_0_1_2_3_c, y_0_1_2_3_d, z_0_1_2_3;
zeros = _mm256_setzero_pd();
/* y_0_1_2_3 = _mm256_setzero_pd(); */
/* y_0_1_2_3_b = _mm256_setzero_pd(); */
/* y_0_1_2_3_c = _mm256_setzero_pd(); */
y_0_1_2_3_d = _mm256_setzero_pd();
// second col (avoid zero y_0_1)
z_0 = _mm_loaddup_pd( &x[1] );
a_00_10 = _mm_load_pd( &A[0+lda*1] );
y_0_1 = _mm_mul_pd( a_00_10, z_0 );
// first col
z_0 = _mm_load_sd( &x[0] );
a_00_10 = _mm_load_sd( &A[0+lda*0] );
tmp0 = _mm_mul_sd( a_00_10, z_0 );
y_0_1 = _mm_add_sd( y_0_1, tmp0 );
y_0_1_2_3_c = _mm256_castpd128_pd256( y_0_1 );
y_0_1_2_3_c = _mm256_blend_pd( y_0_1_2_3_c, y_0_1_2_3_d, 0xc );
// forth col (avoid zero y_0_1_2_3)
x_3 = _mm256_broadcast_sd( &x[3] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
y_0_1_2_3 = _mm256_mul_pd( a_03_13_23_33, x_3 );
// first col
x_2 = _mm256_broadcast_sd( &x[2] );
x_2 = _mm256_blend_pd( x_2, zeros, 0x8 );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
y_0_1_2_3_b = _mm256_mul_pd( a_02_12_22_32, x_2 );
A += 4*lda;
x += 4;
k=4;
for(; k<kmax-3; k+=4)
{
x_0 = _mm256_broadcast_sd( &x[0] );
x_1 = _mm256_broadcast_sd( &x[1] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
x_2 = _mm256_broadcast_sd( &x[2] );
x_3 = _mm256_broadcast_sd( &x[3] );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, ax_temp );
ax_temp = _mm256_mul_pd( a_01_11_21_31, x_1 );
y_0_1_2_3_b = _mm256_add_pd( y_0_1_2_3_b, ax_temp );
ax_temp = _mm256_mul_pd( a_02_12_22_32, x_2 );
y_0_1_2_3_c = _mm256_add_pd( y_0_1_2_3_c, ax_temp );
ax_temp = _mm256_mul_pd( a_03_13_23_33, x_3 );
y_0_1_2_3_d = _mm256_add_pd( y_0_1_2_3_d, ax_temp );
A += 4*lda;
x += 4;
}
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, y_0_1_2_3_c );
y_0_1_2_3_b = _mm256_add_pd( y_0_1_2_3_b, y_0_1_2_3_d );
if(kmax%4>=2)
{
x_0 = _mm256_broadcast_sd( &x[0] );
x_1 = _mm256_broadcast_sd( &x[1] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, ax_temp );
ax_temp = _mm256_mul_pd( a_01_11_21_31, x_1 );
y_0_1_2_3_b = _mm256_add_pd( y_0_1_2_3_b, ax_temp );
A += 2*lda;
x += 2;
}
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, y_0_1_2_3_b );
if(kmax%2==1)
{
x_0 = _mm256_broadcast_sd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, ax_temp );
}
if(alg==0)
{
_mm256_storeu_pd(&y[0], y_0_1_2_3);
}
else if(alg==1)
{
z_0_1_2_3 = _mm256_loadu_pd( &y[0] );
z_0_1_2_3 = _mm256_add_pd ( z_0_1_2_3, y_0_1_2_3 );
_mm256_storeu_pd(&y[0], z_0_1_2_3);
}
else // alg==-1
{
z_0_1_2_3 = _mm256_loadu_pd( &y[0] );
z_0_1_2_3 = _mm256_sub_pd ( z_0_1_2_3, y_0_1_2_3 );
_mm256_storeu_pd(&y[0], z_0_1_2_3);
}
}
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 horizontally inside a block
void kernel_dtrmv_u_n_8_lib4(int kmax, double *A0, int sda, double *x, double *y, int alg)
{
if(kmax<=0)
return;
double *A1 = A0 + 4*sda;
const int lda = 4;
int k;
__m128d
tmp0,
z_0, y_0_1, a_00_10;
__m256d
zeros,
ax_temp,
a_00_10_20_30, a_01_11_21_31,
a_40_50_60_70, a_41_51_61_71,
x_0, x_1,
y_0_1_2_3, y_0_1_2_3_b, z_0_1_2_3,
y_4_5_6_7, y_4_5_6_7_b, z_4_5_6_7;
/* y_0_1_2_3 = _mm256_setzero_pd(); */
/* y_4_5_6_7 = _mm256_setzero_pd(); */
/* y_0_1_2_3_b = _mm256_setzero_pd(); */
/* y_4_5_6_7_b = _mm256_setzero_pd(); */
zeros = _mm256_setzero_pd();
/* y_0_1_2_3 = _mm256_setzero_pd(); */
/* y_0_1_2_3_b = _mm256_setzero_pd(); */
/* y_0_1_2_3_c = _mm256_setzero_pd(); */
/* y_0_1_2_3_d = _mm256_setzero_pd();*/
// upper triangular
// second col (avoid zero y_0_1)
z_0 = _mm_loaddup_pd( &x[1] );
a_00_10 = _mm_load_pd( &A0[0+lda*1] );
y_0_1 = _mm_mul_pd( a_00_10, z_0 );
// first col
z_0 = _mm_load_sd( &x[0] );
a_00_10 = _mm_load_sd( &A0[0+lda*0] );
tmp0 = _mm_mul_sd( a_00_10, z_0 );
y_0_1 = _mm_add_sd( y_0_1, tmp0 );
y_0_1_2_3_b = _mm256_castpd128_pd256( y_0_1 );
y_0_1_2_3_b = _mm256_blend_pd( y_0_1_2_3_b, y_0_1_2_3_b, 0xc );
// forth col (avoid zero y_0_1_2_3)
x_1 = _mm256_broadcast_sd( &x[3] );
a_01_11_21_31 = _mm256_load_pd( &A0[0+lda*3] );
y_0_1_2_3 = _mm256_mul_pd( a_01_11_21_31, x_1 );
// first col
x_0 = _mm256_broadcast_sd( &x[2] );
x_0 = _mm256_blend_pd( x_0, zeros, 0x8 );
a_00_10_20_30 = _mm256_load_pd( &A0[0+lda*2] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3_b = _mm256_add_pd( y_0_1_2_3_b, ax_temp );
A0 += 4*lda;
A1 += 4*lda;
x += 4;
// upper squared
x_0 = _mm256_broadcast_sd( &x[0] );
x_1 = _mm256_broadcast_sd( &x[1] );
a_00_10_20_30 = _mm256_load_pd( &A0[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A0[0+lda*1] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, ax_temp );
ax_temp = _mm256_mul_pd( a_01_11_21_31, x_1 );
y_0_1_2_3_b = _mm256_add_pd( y_0_1_2_3_b, ax_temp );
x_0 = _mm256_broadcast_sd( &x[2] );
x_1 = _mm256_broadcast_sd( &x[3] );
a_00_10_20_30 = _mm256_load_pd( &A0[0+lda*2] );
a_01_11_21_31 = _mm256_load_pd( &A0[0+lda*3] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, ax_temp );
ax_temp = _mm256_mul_pd( a_01_11_21_31, x_1 );
y_0_1_2_3_b = _mm256_add_pd( y_0_1_2_3_b, ax_temp );
// lower triangular
// second col (avoid zero y_0_1)
z_0 = _mm_loaddup_pd( &x[1] );
a_00_10 = _mm_load_pd( &A1[0+lda*1] );
y_0_1 = _mm_mul_pd( a_00_10, z_0 );
// first col
z_0 = _mm_load_sd( &x[0] );
a_00_10 = _mm_load_sd( &A1[0+lda*0] );
tmp0 = _mm_mul_sd( a_00_10, z_0 );
y_0_1 = _mm_add_sd( y_0_1, tmp0 );
y_4_5_6_7_b = _mm256_castpd128_pd256( y_0_1 );
y_4_5_6_7_b = _mm256_blend_pd( y_4_5_6_7_b, y_4_5_6_7_b, 0xc );
// forth col (avoid zero y_4_5_6_7)
x_1 = _mm256_broadcast_sd( &x[3] );
a_01_11_21_31 = _mm256_load_pd( &A1[0+lda*3] );
y_4_5_6_7 = _mm256_mul_pd( a_01_11_21_31, x_1 );
// first col
x_0 = _mm256_broadcast_sd( &x[2] );
x_0 = _mm256_blend_pd( x_0, zeros, 0x8 );
a_00_10_20_30 = _mm256_load_pd( &A1[0+lda*2] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_4_5_6_7_b = _mm256_add_pd( y_4_5_6_7_b, ax_temp );
A0 += 4*lda;
A1 += 4*lda;
x += 4;
k=8;
for(; k<kmax-3; k+=4)
{
/* __builtin_prefetch( A0 + 4*lda );*/
/* __builtin_prefetch( A1 + 4*lda );*/
x_0 = _mm256_broadcast_sd( &x[0] );
x_1 = _mm256_broadcast_sd( &x[1] );
a_00_10_20_30 = _mm256_load_pd( &A0[0+lda*0] );
a_40_50_60_70 = _mm256_load_pd( &A1[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A0[0+lda*1] );
a_41_51_61_71 = _mm256_load_pd( &A1[0+lda*1] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, ax_temp );
ax_temp = _mm256_mul_pd( a_40_50_60_70, x_0 );
y_4_5_6_7 = _mm256_add_pd( y_4_5_6_7, ax_temp );
ax_temp = _mm256_mul_pd( a_01_11_21_31, x_1 );
y_0_1_2_3_b = _mm256_add_pd( y_0_1_2_3_b, ax_temp );
ax_temp = _mm256_mul_pd( a_41_51_61_71, x_1 );
y_4_5_6_7_b = _mm256_add_pd( y_4_5_6_7_b, ax_temp );
/* __builtin_prefetch( A0 + 5*lda );*/
/* __builtin_prefetch( A1 + 5*lda );*/
x_0 = _mm256_broadcast_sd( &x[2] );
x_1 = _mm256_broadcast_sd( &x[3] );
a_00_10_20_30 = _mm256_load_pd( &A0[0+lda*2] );
a_40_50_60_70 = _mm256_load_pd( &A1[0+lda*2] );
a_01_11_21_31 = _mm256_load_pd( &A0[0+lda*3] );
a_41_51_61_71 = _mm256_load_pd( &A1[0+lda*3] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, ax_temp );
ax_temp = _mm256_mul_pd( a_40_50_60_70, x_0 );
y_4_5_6_7 = _mm256_add_pd( y_4_5_6_7, ax_temp );
ax_temp = _mm256_mul_pd( a_01_11_21_31, x_1 );
y_0_1_2_3_b = _mm256_add_pd( y_0_1_2_3_b, ax_temp );
ax_temp = _mm256_mul_pd( a_41_51_61_71, x_1 );
y_4_5_6_7_b = _mm256_add_pd( y_4_5_6_7_b, ax_temp );
A0 += 4*lda;
A1 += 4*lda;
x += 4;
}
if(kmax%4>=2)
{
x_0 = _mm256_broadcast_sd( &x[0] );
x_1 = _mm256_broadcast_sd( &x[1] );
a_00_10_20_30 = _mm256_load_pd( &A0[0+lda*0] );
a_40_50_60_70 = _mm256_load_pd( &A1[0+lda*0] );
a_01_11_21_31 = _mm256_load_pd( &A0[0+lda*1] );
a_41_51_61_71 = _mm256_load_pd( &A1[0+lda*1] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, ax_temp );
ax_temp = _mm256_mul_pd( a_40_50_60_70, x_0 );
y_4_5_6_7 = _mm256_add_pd( y_4_5_6_7, ax_temp );
ax_temp = _mm256_mul_pd( a_01_11_21_31, x_1 );
y_0_1_2_3_b = _mm256_add_pd( y_0_1_2_3_b, ax_temp );
ax_temp = _mm256_mul_pd( a_41_51_61_71, x_1 );
y_4_5_6_7_b = _mm256_add_pd( y_4_5_6_7_b, ax_temp );
A0 += 2*lda;
A1 += 2*lda;
x += 2;
}
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, y_0_1_2_3_b );
y_4_5_6_7 = _mm256_add_pd( y_4_5_6_7, y_4_5_6_7_b );
if(kmax%2==1)
{
x_0 = _mm256_broadcast_sd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A0[0+lda*0] );
a_40_50_60_70 = _mm256_load_pd( &A1[0+lda*0] );
ax_temp = _mm256_mul_pd( a_00_10_20_30, x_0 );
y_0_1_2_3 = _mm256_add_pd( y_0_1_2_3, ax_temp );
ax_temp = _mm256_mul_pd( a_40_50_60_70, x_0 );
y_4_5_6_7 = _mm256_add_pd( y_4_5_6_7, ax_temp );
/* A0 += 1*lda;*/
/* A1 += 1*lda;*/
/* x += 1;*/
}
if(alg==0)
{
_mm256_storeu_pd(&y[0], y_0_1_2_3);
_mm256_storeu_pd(&y[4], y_4_5_6_7);
}
else if(alg==1)
{
z_0_1_2_3 = _mm256_loadu_pd( &y[0] );
z_4_5_6_7 = _mm256_loadu_pd( &y[4] );
z_0_1_2_3 = _mm256_add_pd( z_0_1_2_3, y_0_1_2_3 );
z_4_5_6_7 = _mm256_add_pd( z_4_5_6_7, y_4_5_6_7 );
_mm256_storeu_pd(&y[0], z_0_1_2_3);
_mm256_storeu_pd(&y[4], z_4_5_6_7);
}
else // alg==-1
{
z_0_1_2_3 = _mm256_loadu_pd( &y[0] );
z_4_5_6_7 = _mm256_loadu_pd( &y[4] );
z_0_1_2_3 = _mm256_sub_pd( z_0_1_2_3, y_0_1_2_3 );
z_4_5_6_7 = _mm256_sub_pd( z_4_5_6_7, y_4_5_6_7 );
_mm256_storeu_pd(&y[0], z_0_1_2_3);
_mm256_storeu_pd(&y[4], z_4_5_6_7);
}
}
// 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);
}
}
