void THDoubleVector_copy_AVX(double *y, const double *x, const ptrdiff_t n) {
ptrdiff_t i;
ptrdiff_t off;
for (i=0; i<=((n)-8); i+=8) {
_mm256_storeu_pd(y+i, _mm256_loadu_pd(x+i));
_mm256_storeu_pd(y+i+4, _mm256_loadu_pd(x+i+4));
}
off = (n) - ((n)%8);
for (i=0; i<((n)%8); i++) {
y[off+i] = x[off+i];
}
}
void THDoubleVector_fill_AVX(double *x, const double c, const ptrdiff_t n) {
ptrdiff_t i;
ptrdiff_t off;
__m256d YMM0 = _mm256_set_pd(c, c, c, c);
for (i=0; i<=((n)-16); i+=16) {
_mm256_storeu_pd((x)+i , YMM0);
_mm256_storeu_pd((x)+i+4, YMM0);
_mm256_storeu_pd((x)+i+8, YMM0);
_mm256_storeu_pd((x)+i+12, YMM0);
}
off = (n) - ((n)%16);
for (i=0; i<((n)%16); i++) {
x[off+i] = c;
}
}
// 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;
}
void THDoubleVector_muls_AVX(double *y, const double *x, const double c, const ptrdiff_t n) {
ptrdiff_t i;
__m256d YMM15 = _mm256_set_pd(c, c, c, c);
__m256d YMM0, YMM1;
for (i=0; i<=((n)-8); i+=8) {
YMM0 = _mm256_loadu_pd(x+i);
YMM1 = _mm256_loadu_pd(x+i+4);
YMM0 = _mm256_mul_pd(YMM0, YMM15);
YMM1 = _mm256_mul_pd(YMM1, YMM15);
_mm256_storeu_pd(y+i, YMM0);
_mm256_storeu_pd(y+i+4, YMM1);
}
for (; i<n; i++) {
y[i] = x[i] * c;
}
}
void SimpleClean::PartialSubtractImageAVX(double *image, size_t imgWidth, size_t imgHeight, const double *psf, size_t psfWidth, size_t psfHeight, size_t x, size_t y, double factor, size_t startY, size_t endY)
{
size_t startX, endX;
int offsetX = (int) x - psfWidth/2, offsetY = (int) y - psfHeight/2;
if(offsetX > 0)
startX = offsetX;
else
startX = 0;
if(offsetY > (int) startY)
startY = offsetY;
endX = std::min(x + psfWidth/2, imgWidth);
size_t unAlignedCount = (endX - startX) % 4;
endX -= unAlignedCount;
endY = std::min(y + psfHeight/2, endY);
const __m256d mFactor = _mm256_set1_pd(-factor);
for(size_t ypos = startY; ypos < endY; ++ypos)
{
double *imageIter = image + ypos * imgWidth + startX;
const double *psfIter = psf + (ypos - offsetY) * psfWidth + startX - offsetX;
for(size_t xpos = startX; xpos != endX; xpos+=4)
{
__m256d
imgVal = _mm256_loadu_pd(imageIter),
psfVal = _mm256_loadu_pd(psfIter);
#ifdef __FMA4__
_mm256_storeu_pd(imageIter, _mm256_fmadd_pd(psfVal, mFactor, imgVal));
#else
_mm256_storeu_pd(imageIter, _mm256_add_pd(imgVal, _mm256_mul_pd(psfVal, mFactor)));
#endif
imageIter+=4;
psfIter+=4;
}
for(size_t xpos = endX; xpos!=endX + unAlignedCount; ++xpos)
{
*imageIter -= *psfIter * factor;
++imageIter;
++psfIter;
}
}
}
void THDoubleVector_cmul_AVX(double *z, const double *x, const double *y, const ptrdiff_t n) {
ptrdiff_t i;
__m256d YMM0, YMM1, YMM2, YMM3;
for (i=0; i<=((n)-8); i+=8) {
YMM0 = _mm256_loadu_pd(x+i);
YMM1 = _mm256_loadu_pd(x+i+4);
YMM2 = _mm256_loadu_pd(y+i);
YMM3 = _mm256_loadu_pd(y+i+4);
YMM2 = _mm256_mul_pd(YMM0, YMM2);
YMM3 = _mm256_mul_pd(YMM1, YMM3);
_mm256_storeu_pd(z+i, YMM2);
_mm256_storeu_pd(z+i+4, YMM3);
}
for (; i<n; i++) {
z[i] = x[i] * y[i];
}
}
static inline void matmul_4xkxkx4(int lda, int K, double* a, double* b, double* c)
{
__m256d a_coli, bi0, bi1, bi2, bi3;
__m256d c_col0, c_col1, c_col2, c_col3;
/* layout of 4x4 c matrix
00 01 02 03
10 11 12 13
20 21 22 23
30 31 32 33
*/
double* c01_ptr = c + lda;
double* c02_ptr = c01_ptr + lda;
double* c03_ptr = c02_ptr + lda;
// load old value of c
c_col0 = _mm256_loadu_pd(c);
c_col1 = _mm256_loadu_pd(c01_ptr);
c_col2 = _mm256_loadu_pd(c02_ptr);
c_col3 = _mm256_loadu_pd(c03_ptr);
// for every column of a (or every row of b)
for (int i = 0; i < K; ++i)
{
a_coli = _mm256_load_pd(a);
a += 4;
bi0 = _mm256_broadcast_sd(b++);
bi1 = _mm256_broadcast_sd(b++);
bi2 = _mm256_broadcast_sd(b++);
bi3 = _mm256_broadcast_sd(b++);
c_col0 = _mm256_add_pd(c_col0, _mm256_mul_pd(a_coli, bi0));
c_col1 = _mm256_add_pd(c_col1, _mm256_mul_pd(a_coli, bi1));
c_col2 = _mm256_add_pd(c_col2, _mm256_mul_pd(a_coli, bi2));
c_col3 = _mm256_add_pd(c_col3, _mm256_mul_pd(a_coli, bi3));
}
_mm256_storeu_pd(c, c_col0);
_mm256_storeu_pd(c01_ptr, c_col1);
_mm256_storeu_pd(c02_ptr, c_col2);
_mm256_storeu_pd(c03_ptr, c_col3);
}
void ComplexToDouble(Complex *src, double *dstI, double *dstQ, const unsigned int len)
{
__m128 avxA, avxB;
__m256d avxA_D, avxB_D, avxX_D, avxY_D, avxR_D, avxI_D;
for (unsigned int i=0; i+4<=len; i+=4) {
avxA = _mm_maskload_ps((float*)(src+i), _mm_set_epi32(SET_1, SET_1, SET_1, SET_1)); //load float
avxB = _mm_maskload_ps((float*)(src+i+2), _mm_set_epi32(SET_1, SET_1, SET_1, SET_1));
avxA_D = _mm256_cvtps_pd(avxA); //float to double
avxB_D = _mm256_cvtps_pd(avxB);
avxX_D = _mm256_permute2f128_pd(avxA_D, avxB_D, 0x20);
avxY_D = _mm256_permute2f128_pd(avxA_D, avxB_D, 0x31);
avxR_D = _mm256_shuffle_pd(avxX_D, avxY_D, 0x00);
avxI_D = _mm256_shuffle_pd(avxX_D, avxY_D, 0x0f);
_mm256_storeu_pd(dstI+i, avxR_D); //store
_mm256_storeu_pd(dstQ+i, avxI_D);
}
for (unsigned int i=len-(len&0x03); i<len; ++i) {
dstI[i] = static_cast<double>(src[i].m_real);
dstQ[i] = static_cast<double>(src[i].m_imag);
}
}
extern "C" void product32x32_avx(double *a, double *b, double *c, int n)
{
for(int i=0; i<n; i++) {
__m256d t1 = _mm256_loadu_pd(&c[i*n + 0]);
__m256d t2 = _mm256_loadu_pd(&c[i*n + 4]);
__m256d t3 = _mm256_loadu_pd(&c[i*n + 8]);
__m256d t4 = _mm256_loadu_pd(&c[i*n + 12]);
__m256d t5 = _mm256_loadu_pd(&c[i*n + 16]);
__m256d t6 = _mm256_loadu_pd(&c[i*n + 20]);
__m256d t7 = _mm256_loadu_pd(&c[i*n + 24]);
__m256d t8 = _mm256_loadu_pd(&c[i*n + 28]);
for(int k=0; k<n; k++) {
__m256d a1 = _mm256_set1_pd(a[k*n+i]);
__m256d b1 = _mm256_loadu_pd(&b[k*n+0]);
t1 = _mm256_sub_pd(t1,_mm256_mul_pd(a1,b1));
__m256d b2 = _mm256_loadu_pd(&b[k*n+4]);
t2 = _mm256_sub_pd(t2,_mm256_mul_pd(a1,b2));
__m256d b3 = _mm256_loadu_pd(&b[k*n+8]);
t3 = _mm256_sub_pd(t3,_mm256_mul_pd(a1,b3));
__m256d b4 = _mm256_loadu_pd(&b[k*n+12]);
t4 = _mm256_sub_pd(t4,_mm256_mul_pd(a1,b4));
__m256d b5 = _mm256_loadu_pd(&b[k*n+16]);
t5 = _mm256_sub_pd(t5,_mm256_mul_pd(a1,b5));
__m256d b6 = _mm256_loadu_pd(&b[k*n+20]);
t6 = _mm256_sub_pd(t6,_mm256_mul_pd(a1,b6));
__m256d b7 = _mm256_loadu_pd(&b[k*n+24]);
t7 = _mm256_sub_pd(t7,_mm256_mul_pd(a1,b7));
__m256d b8 = _mm256_loadu_pd(&b[k*n+28]);
t8 = _mm256_sub_pd(t8,_mm256_mul_pd(a1,b8));
}
_mm256_storeu_pd(&c[i*n + 0], t1);
_mm256_storeu_pd(&c[i*n + 4], t2);
_mm256_storeu_pd(&c[i*n + 8], t3);
_mm256_storeu_pd(&c[i*n + 12], t4);
_mm256_storeu_pd(&c[i*n + 16], t5);
_mm256_storeu_pd(&c[i*n + 20], t6);
_mm256_storeu_pd(&c[i*n + 24], t7);
_mm256_storeu_pd(&c[i*n + 28], t8);
}
}
void KERNEL_NAME(VMLLONG n, VML_FLOAT * a, VML_FLOAT * b, VML_FLOAT * y, VML_FLOAT * z, VML_FLOAT * other_params) {
unsigned int m = n >> 2;
unsigned int k = n & 3, j;
unsigned int l = n & (~3);
for (j = 0; j < m; j++) {
v4sd src = _mm256_loadu_pd(a + 4 * j);
v4sd tem = simd_exp4d(src);
_mm256_storeu_pd(y + 4 * j, tem);
}
for (j = 0; j < k; j++) {
y[j + l] = exp(a[j + l]);
}
}
void THDoubleVector_cadd_AVX(double *z, const double *x, const double *y, const double c, const ptrdiff_t n) {
ptrdiff_t i;
__m256d YMM15 = _mm256_set_pd(c, c, c, c);
__m256d YMM0, YMM1, YMM2, YMM3;
for (i=0; i<=((n)-4); i+=4) {
YMM0 = _mm256_loadu_pd(y+i);
YMM1 = _mm256_loadu_pd(x+i);
YMM2 = _mm256_mul_pd(YMM0, YMM15);
YMM3 = _mm256_add_pd(YMM1, YMM2);
_mm256_storeu_pd(z+i, YMM3);
}
for (; i<(n); i++) {
z[i] = x[i] + y[i] * c;
}
}
void static
avx_test (void)
{
int i;
union256d u, s1, s2;
long long source1[4]={34545, 95567, 23443, 5675};
long long source2[4]={674, 57897, 93459, 45624};
long long d[4];
long long e[4];
s1.x = _mm256_loadu_pd ((double *)source1);
s2.x = _mm256_loadu_pd ((double *)source2);
u.x = _mm256_andnot_pd (s1.x, s2.x);
_mm256_storeu_pd ((double *)d, u.x);
for (i = 0; i < 4; i++)
e[i] = (~source1[i]) & source2[i];
if (checkVl (d, e, 4))
abort ();
}
/* sum double vectors ----------------------------------------------------------
* sum double vectors: out=data1.+data2
* args : double *data1 I input double array
* double *data2 I input double array
* int n I number of input data
* double *out O output double array
* return : none
* note : AVX command is used if "AVX" is defined
*-----------------------------------------------------------------------------*/
extern void sumvd(const double *data1, const double *data2, int n, double *out)
{
int i;
#if !defined(AVX_ENABLE)
for (i=0;i<n;i++) out[i]=data1[i]+data2[i];
#else
int m=n/4;
__m256d xmm1,xmm2,xmm3;
if (n<8) {
for (i=0;i<n;i++) out[i]=data1[i]+data2[i];
}
else {
for (i=0;i<4*m;i+=4) {
xmm1=_mm256_loadu_pd(&data1[i]);
xmm2=_mm256_loadu_pd(&data2[i]);
xmm3=_mm256_add_pd(xmm1,xmm2);
_mm256_storeu_pd(&out[i],xmm3);
}
for (;i<n;i++) out[i]=data1[i]+data2[i];
}
#endif
}
void jacobi_avx(GRID_T *oldGrid, GRID_T *newGrid, int width, int height){
int remainder;
remainder = (width-2)%4;
/* Each vector contains one value of the four Jacobi iteration step
* Either each upper, below, left or right value. */
__m256d up_row, below_row, right_row, left_row;
__m256d factor = _mm256_set1_pd(0.25);
for(int i = 1; i < height-1; i++){
for(int j = 1; j < width-4; j += 4){
up_row = _mm256_loadu_pd(&(oldGrid[(i-1)*width + j]));
below_row = _mm256_loadu_pd(&(oldGrid[(i+1)*width + j]));
right_row = _mm256_loadu_pd(&(oldGrid[i*width + (j+1)]));
left_row = _mm256_loadu_pd(&(oldGrid[i*width + (j-1)]));
/* Sum up n-th element of each vector */
__m256d dest;
__m256d add_1 = _mm256_add_pd(up_row, below_row);
__m256d add_2 = _mm256_add_pd(left_row, right_row);
dest = _mm256_add_pd(add_2, add_1);
/* Multiplicat with 0.25 */
dest = _mm256_mul_pd(dest, factor);
// Use unaligned store method. Normal one produces segmentation fault
_mm256_storeu_pd(&(newGrid[i*width + j]), dest);
}
for(int j = width - remainder - 1; j < width -1; j++){
newGrid[i*width + j] = (oldGrid[i*width + (j-1)] + oldGrid[i*width + (j+1)] + oldGrid[(i-1)*width + j] + oldGrid[(i+1)*width + j]) * 0.25;
}
}
return;
}
// 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);
}
}
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);
}
}
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);
}
}
// it moves horizontally inside a block
void kernel_dgemv_n_8_lib4(int kmax, double *A0, double *A1, double *x, double *y, int alg)
{
if(kmax<=0)
return;
const int lda = 4;
int k;
__m256d
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();
if(kmax<=64)
{
k=0;
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;
}
}
else
{
k=0;
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 );*/
__builtin_prefetch( A0 + 6*lda );
__builtin_prefetch( A1 + 6*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);
}
}
/*!
* \brief Unaligned store of the given packed vector at the
* given memory position
*/
ETL_STATIC_INLINE(void) storeu(etl::complex<double>* memory, avx_simd_complex_double<etl::complex<double>> value) {
_mm256_storeu_pd(reinterpret_cast<double*>(memory), value.value);
}
/*!
* \brief Unaligned store of the given packed vector at the
* given memory position
*/
ETL_STATIC_INLINE(void) storeu(double* memory, avx_simd_double value) {
_mm256_storeu_pd(memory, value.value);
}
// it moves horizontally inside a block
void kernel_dgemv_n_4_lib4(int kmax, double *A, double *x, double *y, int alg)
{
if(kmax<=0)
return;
const int lda = 4;
int k;
__m256d
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;
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();
k=0;
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 );
/* A += 1*lda;*/
/* x += 1;*/
}
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 vector4d::store_unaligned(double* dst) const
{
_mm256_storeu_pd(dst, m_value);
}
// 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 );
}
}
test (double *e, __m256d a)
{
return _mm256_storeu_pd (e, a);
}
KFR_INTRINSIC void write(cunaligned_t, f64* ptr, const f64avx& x) { _mm256_storeu_pd(ptr, x.v); }
// it moves vertically across blocks
void kernel_dtrmv_u_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;
__m256d
zeros,
tmp0, tmp1,
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_00 = _mm256_setzero_pd();
y_11 = _mm256_setzero_pd();
y_22 = _mm256_setzero_pd();
y_33 = _mm256_setzero_pd();
k=0;
for(; k<kmax-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] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, tmp0 );
y_11 = _mm256_add_pd( y_11, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, tmp0 );
y_33 = _mm256_add_pd( y_33, tmp1 );
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] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, tmp0 );
y_11 = _mm256_add_pd( y_11, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, tmp0 );
y_33 = _mm256_add_pd( y_33, tmp1 );
A += 4 + (sda-1)*lda;
x += 4;
}
for(; k<kmax-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] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, tmp0 );
y_11 = _mm256_add_pd( y_11, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, tmp0 );
y_33 = _mm256_add_pd( y_33, tmp1 );
A += 4 + (sda-1)*lda;
x += 4;
}
zeros = _mm256_setzero_pd();
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_00_10_20_30 = _mm256_blend_pd( a_00_10_20_30, zeros, 0xe );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
a_01_11_21_31 = _mm256_blend_pd( a_01_11_21_31, zeros, 0xc );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
a_02_12_22_32 = _mm256_blend_pd( a_02_12_22_32, zeros, 0x8 );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, tmp0 );
y_11 = _mm256_add_pd( y_11, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, tmp0 );
y_33 = _mm256_add_pd( y_33, tmp1 );
__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);
}
}
// it moves vertically across blocks
void kernel_dtrmv_u_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-kna; // number from aligned positon*/
__m256d
zeros,
tmp0, tmp1,
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();
k=0;
for(; k<kmax-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] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, tmp0 );
y_11 = _mm256_add_pd( y_11, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, tmp0 );
y_33 = _mm256_add_pd( y_33, tmp1 );
/* __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] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_44 = _mm256_add_pd( y_44, tmp0 );
y_55 = _mm256_add_pd( y_55, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_66 = _mm256_add_pd( y_66, tmp0 );
y_77 = _mm256_add_pd( y_77, tmp1 );
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] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, tmp0 );
y_11 = _mm256_add_pd( y_11, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, tmp0 );
y_33 = _mm256_add_pd( y_33, tmp1 );
/* __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] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_44 = _mm256_add_pd( y_44, tmp0 );
y_55 = _mm256_add_pd( y_55, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_66 = _mm256_add_pd( y_66, tmp0 );
y_77 = _mm256_add_pd( y_77, tmp1 );
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;*/
/* }*/
zeros = _mm256_setzero_pd();
// top triangle
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*0] );
a_00_10_20_30 = _mm256_blend_pd( a_00_10_20_30, zeros, 0xe );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*1] );
a_01_11_21_31 = _mm256_blend_pd( a_01_11_21_31, zeros, 0xc );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*2] );
a_02_12_22_32 = _mm256_blend_pd( a_02_12_22_32, zeros, 0x8 );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*3] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_00 = _mm256_add_pd( y_00, tmp0 );
y_11 = _mm256_add_pd( y_11, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_22 = _mm256_add_pd( y_22, tmp0 );
y_33 = _mm256_add_pd( y_33, tmp1 );
// top square
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] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_44 = _mm256_add_pd( y_44, tmp0 );
y_55 = _mm256_add_pd( y_55, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_66 = _mm256_add_pd( y_66, tmp0 );
y_77 = _mm256_add_pd( y_77, tmp1 );
A += 4 + (sda-1)*lda;
x += 4;
// bottom triangle
x_0_1_2_3 = _mm256_loadu_pd( &x[0] );
a_00_10_20_30 = _mm256_load_pd( &A[0+lda*4] );
a_00_10_20_30 = _mm256_blend_pd( a_00_10_20_30, zeros, 0xe );
a_01_11_21_31 = _mm256_load_pd( &A[0+lda*5] );
a_01_11_21_31 = _mm256_blend_pd( a_01_11_21_31, zeros, 0xc );
a_02_12_22_32 = _mm256_load_pd( &A[0+lda*6] );
a_02_12_22_32 = _mm256_blend_pd( a_02_12_22_32, zeros, 0x8 );
a_03_13_23_33 = _mm256_load_pd( &A[0+lda*7] );
tmp0 = _mm256_mul_pd( a_00_10_20_30, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_01_11_21_31, x_0_1_2_3 );
y_44 = _mm256_add_pd( y_44, tmp0 );
y_55 = _mm256_add_pd( y_55, tmp1 );
tmp0 = _mm256_mul_pd( a_02_12_22_32, x_0_1_2_3 );
tmp1 = _mm256_mul_pd( a_03_13_23_33, x_0_1_2_3 );
y_66 = _mm256_add_pd( y_66, tmp0 );
y_77 = _mm256_add_pd( y_77, tmp1 );
// store
__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);
}
}
// 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);
}
}
ALGEBRA_INLINE void vector_addm_double_aligned_32 (double* v1,double lambda,const double* v2,size_t n)
{
size_t k;
__m256d l1 = _mm256_broadcast_sd(&lambda);
__m256d l2 = _mm256_broadcast_sd(&lambda);
__m256d l3 = _mm256_broadcast_sd(&lambda);
__m256d l4 = _mm256_broadcast_sd(&lambda);
size_t q = n / 16;
size_t r = n % 16;
if(q > 0) {
if (ALGEBRA_IS_ALIGNED(v1) && ALGEBRA_IS_ALIGNED(v2)) {
for (k=0;k<q;k++) {
/* Charge 4 valeurs de chaque tableau */
__m256d i1 = _mm256_load_pd(v1);
__m256d j1 = _mm256_load_pd(v2);
__m256d i2 = _mm256_load_pd(v1+4);
__m256d j2 = _mm256_load_pd(v2+4);
__m256d i3 = _mm256_load_pd(v1+8);
__m256d j3 = _mm256_load_pd(v2+8);
__m256d i4 = _mm256_load_pd(v1+12);
__m256d j4 = _mm256_load_pd(v2+12);
/* multiplie */
j1 = _mm256_mul_pd(j1, l1);
j2 = _mm256_mul_pd(j2, l2);
j3 = _mm256_mul_pd(j3, l3);
j4 = _mm256_mul_pd(j4, l4);
/* Additionne */
i1 = _mm256_add_pd(i1,j1);
i2 = _mm256_add_pd(i2,j2);
i3 = _mm256_add_pd(i3,j3);
i4 = _mm256_add_pd(i4,j4);
/* Sauvegarde */
_mm256_store_pd(v1, i1);
_mm256_store_pd(v1+4, i2);
_mm256_store_pd(v1+8, i3);
_mm256_store_pd(v1+12, i4);
v1 += 16;
v2 += 16;
}
}
else {
for (k=0;k<q;k++) {
/* Charge 4 valeurs de chaque tableau */
__m256d i1 = _mm256_loadu_pd(v1);
__m256d j1 = _mm256_loadu_pd(v2);
__m256d i2 = _mm256_loadu_pd(v1+4);
__m256d j2 = _mm256_loadu_pd(v2+4);
__m256d i3 = _mm256_loadu_pd(v1+8);
__m256d j3 = _mm256_loadu_pd(v2+8);
__m256d i4 = _mm256_loadu_pd(v1+12);
__m256d j4 = _mm256_loadu_pd(v2+12);
/* multiplie */
j1 = _mm256_mul_pd(j1, l1);
j2 = _mm256_mul_pd(j2, l2);
j3 = _mm256_mul_pd(j3, l3);
j4 = _mm256_mul_pd(j4, l4);
/* Additionne */
i1 = _mm256_add_pd(i1,j1);
i2 = _mm256_add_pd(i2,j2);
i3 = _mm256_add_pd(i3,j3);
i4 = _mm256_add_pd(i4,j4);
/* Sauvegarde */
_mm256_storeu_pd(v1, i1);
_mm256_storeu_pd(v1+4, i2);
_mm256_storeu_pd(v1+8, i3);
_mm256_storeu_pd(v1+12, i4);
v1 += 16;
v2 += 16;
}
}
}
for(k = 0 ; k<r ; k++)
v1[k] += lambda*v2[k];
}
// add *p by *s and applied to all n
COREARRAY_DLL_DEFAULT void vec_f64_add(double *p, const double *s, size_t n)
{
#if defined(COREARRAY_SIMD_AVX)
switch ((size_t)p & 0x1F)
{
case 0x08:
if (n > 0) { (*p++) += (*s++); n--; }
case 0x10:
if (n > 0) { (*p++) += (*s++); n--; }
case 0x18:
if (n > 0) { (*p++) += (*s++); n--; }
case 0x00:
for (; n >= 4; n-=4)
{
_mm256_store_pd(p, _mm256_add_pd(_mm256_load_pd(p), _mm256_loadu_pd(s)));
p += 4; s += 4;
}
if (n >= 2)
{
_mm_store_pd(p, _mm_add_pd(_mm_load_pd(p), _mm_loadu_pd(s)));
p += 2; s += 2; n -= 2;
}
break;
default:
for (; n >= 4; n-=4)
{
_mm256_storeu_pd(p, _mm256_add_pd(_mm256_loadu_pd(p), _mm256_loadu_pd(s)));
p += 4; s += 4;
}
if (n >= 2)
{
_mm_storeu_pd(p, _mm_add_pd(_mm_loadu_pd(p), _mm_loadu_pd(s)));
p += 2; s += 2; n -= 2;
}
}
#elif defined(COREARRAY_SIMD_SSE2)
switch ((size_t)p & 0x0F)
{
case 0x08:
if (n > 0) { (*p++) += (*s++); n--; }
case 0x00:
for (; n >= 2; n-=2)
{
_mm_store_pd(p, _mm_add_pd(_mm_load_pd(p), _mm_loadu_pd(s)));
p += 2; s += 2;
}
break;
default:
for (; n >= 2; n-=2)
{
_mm_storeu_pd(p, _mm_add_pd(_mm_loadu_pd(p), _mm_loadu_pd(s)));
p += 2; s += 2;
}
}
#endif
for (; n > 0; n--) (*p++) += (*s++);
}
