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 static
avx_test (void)
{
int i;
long long m[4] = {mask_v(0), mask_v(1), mask_v(2), mask_v(3)};
double s[4] = {1.1, 2.2, 3.3, 4.4};
union256d u;
union256i_q mask;
double e [4] = {0.0};
mask.x = _mm256_loadu_si256 ((__m256i *)m);
u.x = _mm256_maskload_pd (s, mask.x);
for (i = 0 ; i < 4; i++)
e[i] = m[i] ? s[i] : 0;
if (check_union256d (u, e))
abort ();
}
/**
* Calculate all values in one step per pixel. Requires grabbing the neighboring pixels.
*/
FORCE_INLINE double single_pixel(
double *im, int center, int top, int left, int right, int bottom,
const __m256i mask1110,
const __m256d rgb0W,
const __m256d onehalf,
const __m256d minustwelvehalf){
// double r = im[center];
// double g = im[center+1];
// double b = im[center+2];
// double r1 = im[top];
// double g1 = im[top+1];
// double b1 = im[top+2];
// double r2 = im[left];
// double g2 = im[left+1];
// double b2 = im[left+2];
// double r3 = im[right];
// double g3 = im[right+1];
// double b3 = im[right+2];
// double r4 = im[bottom];
// double g4 = im[bottom+1];
// double b4 = im[bottom+2];
__m256d c = _mm256_maskload_pd(&(im[center]),mask1110);
__m256d c1 = _mm256_loadu_pd(&(im[top]));
__m256d c2 = _mm256_loadu_pd(&(im[left]));
__m256d c3 = _mm256_loadu_pd(&(im[right]));
__m256d c4 = _mm256_loadu_pd(&(im[bottom]));
COST_INC_LOAD(20);
// double grey = rw * r + gw * g + bw * b;
// double grey1 = rw * r1 + gw * g1 + bw * b1;
// double grey2 = rw * r2 + gw * g2 + bw * b2;
// double grey3 = rw * r3 + gw * g3 + bw * b3;
// double grey4 = rw * r4 + gw * g4 + bw * b4;
__m256d greyc = _mm256_mul_pd(c,rgb0W);
__m256d grey1 = _mm256_mul_pd(c1,rgb0W);
__m256d grey2 = _mm256_mul_pd(c2,rgb0W);
__m256d grey3 = _mm256_mul_pd(c3,rgb0W);
__m256d grey4 = _mm256_mul_pd(c4,rgb0W);
//AVX: double: horizontal add for 1 vector
__m256d c_perm = _mm256_permute2f128_pd(c, c, 0b00100001);//1,2
__m256d c_h = _mm256_hadd_pd(c,c_perm);
__m128d c_h_lo = _mm256_extractf128_pd (c_h, 0);// lo
__m128d c_h_hi = _mm256_extractf128_pd (c_h, 1);// hi
double c_hsum_lo = _mm_cvtsd_f64(c_h_lo);
double c_hsum_hi = _mm_cvtsd_f64(c_h_hi);
double c_hsum = c_hsum_lo + c_hsum_hi;
//AVX: double: horizontal add for 1 vector
__m256d greyc_perm = _mm256_permute2f128_pd(greyc, greyc, 0b00100001);//1,2
__m256d greyc_h = _mm256_hadd_pd(greyc,greyc_perm);
__m128d greyc_h_lo = _mm256_extractf128_pd (greyc_h, 0);// lo
__m128d greyc_h_hi = _mm256_extractf128_pd (greyc_h, 1);// hi
double greyc_hsum_lo = _mm_cvtsd_f64(greyc_h_lo);
double greyc_hsum_hi = _mm_cvtsd_f64(greyc_h_hi);
double greyc_hsum = greyc_hsum_lo + greyc_hsum_hi;
//AVX: _m256d: horizontal add for 4 vectors at once
__m256d grey12 = _mm256_hadd_pd(grey1,grey2);
__m256d grey34 = _mm256_hadd_pd(grey3,grey4);
__m256d grey_1234_blend = _mm256_blend_pd(grey12, grey34, 0b1100); //0011
__m256d grey_1234_perm = _mm256_permute2f128_pd(grey12, grey34, 0b00100001);//1,2
__m256d grey_1234 = _mm256_add_pd(grey_1234_perm, grey_1234_blend);
//AVX: double: horizontal add for 1 vector
__m256d grey1234_perm = _mm256_permute2f128_pd(grey_1234, grey_1234, 0b00100001);//1,2
__m256d grey1234_h = _mm256_hadd_pd(grey_1234,grey1234_perm);
__m128d grey1234_h_lo = _mm256_extractf128_pd (grey1234_h, 0);// lo
__m128d grey1234_h_hi = _mm256_extractf128_pd (grey1234_h, 1);// hi
double grey1234_hsum_lo = _mm_cvtsd_f64(grey1234_h_lo);
double grey1234_hsum_hi = _mm_cvtsd_f64(grey1234_h_hi);
double grey1234_sum = grey1234_hsum_lo + grey1234_hsum_hi;
COST_INC_ADD(10); //+ operations wasted on AVX
COST_INC_MUL(15); //+ operations wasted on AVX
double mu = c_hsum / 3.0;
COST_INC_ADD(2);
COST_INC_DIV(1);
// double rmu = r-mu;
// double gmu = g-mu;
// double bmu = b-mu;
__m256d c_mu = _mm256_set1_pd(mu);
__m256d c_rgbmu = _mm256_sub_pd(c,c_mu);
COST_INC_ADD(3); //+1 operations wasted on AVX
// double rz = r-0.5;
// double gz = g-0.5;
// double bz = b-0.5;
__m256d c_rgbz = _mm256_sub_pd(c,onehalf);
COST_INC_ADD(3); //+1 operations wasted on AVX
// double rzrz = rz*rz;
// double gzgz = gz*gz;
// double bzbz = bz*bz;
__m256d c_rgbz_sq = _mm256_mul_pd(c_rgbz,c_rgbz);
COST_INC_MUL(3); //+1 operations wasted on AVX
// double re = exp(-12.5*rzrz);
// double ge = exp(-12.5*gzgz);
// double be = exp(-12.5*bzbz);
__m256d c_rgbe_tmp = _mm256_mul_pd(minustwelvehalf,c_rgbz_sq);
__m128 c_rgbe_tmp_ps = _mm256_cvtpd_ps(c_rgbe_tmp);
__m128 c_rgbe_ps = exp_ps(c_rgbe_tmp_ps);
__m256d c_rgbe = _mm256_cvtps_pd(c_rgbe_ps);
COST_INC_EXP(3);
COST_INC_MUL(3); //+1 operations wasted on AVX
// double t1 = sqrt((rmu*rmu + gmu*gmu + bmu*bmu)/3.0);
__m256d c_rgbmu_sq = _mm256_mul_pd(c_rgbmu,c_rgbmu);
__m128d t1_tmp1_lo = _mm256_extractf128_pd (c_rgbmu_sq, 0);// lo
__m128d t1_tmp1_hi = _mm256_extractf128_pd (c_rgbmu_sq, 1);// hi
__m128d t1_tmp1_lo_sum = _mm_hadd_pd (t1_tmp1_lo, t1_tmp1_lo);
double t1_tmp1_hi_lo = _mm_cvtsd_f64(t1_tmp1_hi);
double t1_tmp1_lo_sum_lo = _mm_cvtsd_f64(t1_tmp1_lo_sum);
double t1_tmp1 = t1_tmp1_lo_sum_lo + t1_tmp1_hi_lo;
double t1_tmp2 = t1_tmp1 / 3.0;
double t1 = sqrt(t1_tmp2);
COST_INC_SQRT(1);
COST_INC_ADD(3);
COST_INC_MUL(3); //+1 operations wasted on AVX
COST_INC_DIV(1);
double t2 = fabs(t1);
COST_INC_ABS(1);
// double t3 = re*ge*be;
__m128d t3_tmp1_lo = _mm256_extractf128_pd (c_rgbe, 0);// lo
__m128d t3_tmp1_hi = _mm256_extractf128_pd (c_rgbe, 1);// hi
double t3_tmp1_lo_lo = _mm_cvtsd_f64(t3_tmp1_lo);
double t3_tmp1_hi_lo = _mm_cvtsd_f64(t3_tmp1_hi);
__m128d t3_tmp1_lo_swapped = _mm_permute_pd(t3_tmp1_lo, 1);// swap
double t3_tmp1_lo_hi = _mm_cvtsd_f64(t3_tmp1_lo_swapped);
double t3 = t3_tmp1_lo_lo * t3_tmp1_lo_hi * t3_tmp1_hi_lo;
COST_INC_MUL(2);
double t4 = fabs(t3);
COST_INC_ABS(1);
double t5 = t2 * t4;
COST_INC_MUL(1);
// double t6 = -4.0*grey+grey1+grey2+grey3+grey4;
double minusfour_times_grey = -4.0*greyc_hsum;
double t6 = minusfour_times_grey+grey1234_sum;
COST_INC_MUL(1);
COST_INC_ADD(2); //2 operations saved due to AVX
double t7 = fabs(t6);
COST_INC_ABS(1);
double t8 = t5 * t7;
COST_INC_MUL(1);
double t9 = t8 + 1.0E-12;
COST_INC_ADD(1);
return t9;
}
// 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 );
}
}
