Gitiles
Code Review Sign In
realtimeroboticsgroup.org / RealtimeRoboticsGroup / test / 9a24b370cbc07492fa6c205f1027367fa55b6742 / . / blas / d_lapack_lib4.c
blob: 75a4a4fd82df5d71e406eaa996dff639a1adaf66 [file] [log] [blame]
/**************************************************************************************************
* *
* This file is part of BLASFEO. *
* *
* BLASFEO -- BLAS For Embedded Optimization. *
* Copyright (C) 2016-2017 by Gianluca Frison. *
* Developed at IMTEK (University of Freiburg) under the supervision of Moritz Diehl. *
* All rights reserved. *
* *
* HPMPC is free software; you can redistribute it and/or *
* modify it under the terms of the GNU Lesser General Public *
* License as published by the Free Software Foundation; either *
* version 2.1 of the License, or (at your option) any later version. *
* *
* HPMPC is distributed in the hope that it will be useful, *
* but WITHOUT ANY WARRANTY; without even the implied warranty of *
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. *
* See the GNU Lesser General Public License for more details. *
* *
* You should have received a copy of the GNU Lesser General Public *
* License along with HPMPC; if not, write to the Free Software *
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA *
* *
* Author: Gianluca Frison, giaf (at) dtu.dk *
* gianluca.frison (at) imtek.uni-freiburg.de *
* *
**************************************************************************************************/
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include "../include/blasfeo_common.h"
#include "../include/blasfeo_d_aux.h"
#include "../include/blasfeo_d_kernel.h"
/****************************
* old interface
****************************/
void dgetrf_nn_nopivot_lib(int m, int n, double *pC, int sdc, double *pD, int sdd, double *inv_diag_D)
{
if(m<=0 || n<=0)
return;
const int ps = 4;
int ii, jj, ie;
// main loop
ii = 0;
#if defined(TARGET_X64_INTEL_HASWELL)
for( ; ii<m-11; ii+=12)
{
jj = 0;
// solve lower
ie = n<ii ? n : ii; // ie is multiple of 4
for( ; jj<ie-3; jj+=4)
{
kernel_dtrsm_nn_ru_inv_12x4_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+jj*sdd], &inv_diag_D[jj]);
}
if(jj<ie)
{
kernel_dtrsm_nn_ru_inv_12x4_vs_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+jj*sdd], &inv_diag_D[jj], m-ii, ie-jj);
jj+=4;
}
// factorize
if(jj<n-3)
{
kernel_dgetrf_nn_l_12x4_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj]);
jj+=4;
}
else if(jj<n)
{
kernel_dgetrf_nn_l_12x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj], m-ii, n-jj);
jj+=4;
}
if(jj<n-3)
{
kernel_dgetrf_nn_m_12x4_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj]);
jj+=4;
}
else if(jj<n)
{
kernel_dgetrf_nn_m_12x4_vs_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj], m-ii, n-jj);
jj+=4;
}
if(jj<n-3)
{
kernel_dgetrf_nn_r_12x4_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj]);
jj+=4;
}
else if(jj<n)
{
kernel_dgetrf_nn_r_12x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj], m-ii, n-jj);
jj+=4;
}
// solve upper
for( ; jj<n-3; jj+=4)
{
kernel_dtrsm_nn_ll_one_12x4_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[ii*ps+ii*sdd], sdd);
}
if(jj<n)
{
kernel_dtrsm_nn_ll_one_12x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[ii*ps+ii*sdd], sdd, m-ii, n-jj);
}
}
if(m>ii)
{
if(m-ii<=4)
{
goto left_4;
}
else if(m-ii<=8)
{
goto left_8;
}
else
{
goto left_12;
}
}
#elif defined(TARGET_X64_INTEL_SANDY_BRIDGE)
for( ; ii<m-7; ii+=8)
{
jj = 0;
// solve lower
ie = n<ii ? n : ii; // ie is multiple of 4
for( ; jj<ie-3; jj+=4)
{
kernel_dtrsm_nn_ru_inv_8x4_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+jj*sdd], &inv_diag_D[jj]);
}
if(jj<ie)
{
kernel_dtrsm_nn_ru_inv_8x4_vs_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+jj*sdd], &inv_diag_D[jj], m-ii, ie-jj);
jj+=4;
}
// factorize
if(jj<n-3)
{
kernel_dgetrf_nn_l_8x4_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj]);
// kernel_dgetrf_nn_4x4_lib4(jj, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &inv_diag_D[jj]);
// kernel_dtrsm_nn_ru_inv_4x4_lib4(jj, &pD[(ii+4)*sdd], &pD[jj*ps], sdd, &pC[jj*ps+(ii+4)*sdc], &pD[jj*ps+(ii+4)*sdd], &pD[jj*ps+jj*sdd], &inv_diag_D[jj]);
jj+=4;
}
else if(jj<n)
{
kernel_dgetrf_nn_l_8x4_vs_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj], m-ii, n-jj);
// kernel_dgetrf_nn_4x4_vs_lib4(jj, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &inv_diag_D[jj], m-ii, n-jj);
// kernel_dtrsm_nn_ru_inv_4x4_vs_lib4(jj, &pD[(ii+4)*sdd], &pD[jj*ps], sdd, &pC[jj*ps+(ii+4)*sdc], &pD[jj*ps+(ii+4)*sdd], &pD[jj*ps+jj*sdd], &inv_diag_D[jj], m-(ii+4), n-jj);
jj+=4;
}
if(jj<n-3)
{
kernel_dtrsm_nn_ll_one_4x4_lib4(ii, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &pD[ii*ps+ii*sdd]);
kernel_dgetrf_nn_4x4_lib4(jj, &pD[(ii+4)*sdd], &pD[jj*ps], sdd, &pC[jj*ps+(ii+4)*sdc], &pD[jj*ps+(ii+4)*sdd], &inv_diag_D[jj]);
jj+=4;
}
else if(jj<n)
{
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(ii, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &pD[ii*ps+ii*sdd], m-ii, n-jj);
kernel_dgetrf_nn_4x4_vs_lib4(jj, &pD[(ii+4)*sdd], &pD[jj*ps], sdd, &pC[jj*ps+(ii+4)*sdc], &pD[jj*ps+(ii+4)*sdd], &inv_diag_D[jj], m-(ii+4), n-jj);
jj+=4;
}
// solve upper
for( ; jj<n-3; jj+=4)
{
kernel_dtrsm_nn_ll_one_8x4_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd],sdd, &pD[ii*ps+ii*sdd], sdd);
}
if(jj<n)
{
kernel_dtrsm_nn_ll_one_8x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[ii*ps+ii*sdd], sdd, m-ii, n-jj);
}
}
if(m>ii)
{
if(m-ii<=4)
{
goto left_4;
}
else
{
goto left_8;
}
}
#else
for( ; ii<m-3; ii+=4)
{
jj = 0;
// solve lower
ie = n<ii ? n : ii; // ie is multiple of 4
for( ; jj<ie-3; jj+=4)
{
kernel_dtrsm_nn_ru_inv_4x4_lib4(jj, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &pD[jj*ps+jj*sdd], &inv_diag_D[jj]);
}
if(jj<ie)
{
kernel_dtrsm_nn_ru_inv_4x4_vs_lib4(jj, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &pD[jj*ps+jj*sdd], &inv_diag_D[jj], m-ii, ie-jj);
jj+=4;
}
// factorize
if(jj<n-3)
{
kernel_dgetrf_nn_4x4_lib4(jj, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &inv_diag_D[jj]);
jj+=4;
}
else if(jj<n)
{
kernel_dgetrf_nn_4x4_vs_lib4(jj, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &inv_diag_D[jj], m-ii, n-jj);
jj+=4;
}
// solve upper
for( ; jj<n-3; jj+=4)
{
kernel_dtrsm_nn_ll_one_4x4_lib4(ii, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &pD[ii*ps+ii*sdd]);
}
if(jj<n)
{
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(ii, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &pD[ii*ps+ii*sdd], m-ii, n-jj);
}
}
if(m>ii)
{
goto left_4;
}
#endif
// common return if i==m
return;
#if defined(TARGET_X64_INTEL_HASWELL)
left_12:
jj = 0;
// solve lower
ie = n<ii ? n : ii; // ie is multiple of 4
for( ; jj<ie; jj+=4)
{
kernel_dtrsm_nn_ru_inv_12x4_vs_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+jj*sdd], &inv_diag_D[jj], m-ii, ie-jj);
}
// factorize
if(jj<n)
{
kernel_dgetrf_nn_l_12x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj], m-ii, n-jj);
jj+=4;
}
if(jj<n)
{
kernel_dgetrf_nn_l_12x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj], m-ii, n-jj);
jj+=4;
}
if(jj<n)
{
kernel_dgetrf_nn_r_12x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj], m-ii, n-jj);
jj+=4;
}
// solve upper
for( ; jj<n; jj+=4)
{
kernel_dtrsm_nn_ll_one_12x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[ii*ps+ii*sdd], sdd, m-ii, n-jj);
}
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL) || defined(TARGET_X64_INTEL_SANDY_BRIDGE)
left_8:
jj = 0;
// solve lower
ie = n<ii ? n : ii; // ie is multiple of 4
for( ; jj<ie; jj+=4)
{
kernel_dtrsm_nn_ru_inv_8x4_vs_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+jj*sdd], &inv_diag_D[jj], m-ii, ie-jj);
}
// factorize
if(jj<n)
{
kernel_dgetrf_nn_l_8x4_vs_lib4(jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &inv_diag_D[jj], m-ii, n-jj);
// kernel_dgetrf_nn_4x4_vs_lib4(jj, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &inv_diag_D[jj], m-ii, n-jj);
// kernel_dtrsm_nn_ru_inv_4x4_vs_lib4(jj, &pD[(ii+4)*sdd], &pD[jj*ps], sdd, &pC[jj*ps+(ii+4)*sdc], &pD[jj*ps+(ii+4)*sdd], &pD[jj*ps+jj*sdd], &inv_diag_D[jj], m-(ii+4), n-jj);
jj+=4;
}
if(jj<n)
{
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(ii, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &pD[ii*ps+ii*sdd], m-ii, n-jj);
kernel_dgetrf_nn_4x4_vs_lib4(jj, &pD[(ii+4)*sdd], &pD[jj*ps], sdd, &pC[jj*ps+(ii+4)*sdc], &pD[jj*ps+(ii+4)*sdd], &inv_diag_D[jj], m-(ii+4), n-jj);
jj+=4;
}
// solve upper
for( ; jj<n; jj+=4)
{
kernel_dtrsm_nn_ll_one_8x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], sdc, &pD[jj*ps+ii*sdd], sdd, &pD[ii*ps+ii*sdd], sdd, m-ii, n-jj);
}
return;
#endif
left_4:
jj = 0;
// solve lower
ie = n<ii ? n : ii; // ie is multiple of 4
for( ; jj<ie; jj+=4)
{
kernel_dtrsm_nn_ru_inv_4x4_vs_lib4(jj, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &pD[jj*ps+jj*sdd], &inv_diag_D[jj], m-ii, ie-jj);
}
// factorize
if(jj<n)
{
kernel_dgetrf_nn_4x4_vs_lib4(jj, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &inv_diag_D[jj], m-ii, n-jj);
jj+=4;
}
// solve upper
for( ; jj<n; jj+=4)
{
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(ii, &pD[ii*sdd], &pD[jj*ps], sdd, &pC[jj*ps+ii*sdc], &pD[jj*ps+ii*sdd], &pD[ii*ps+ii*sdd], m-ii, n-jj);
}
return;
}
void dgetrf_nn_lib(int m, int n, double *pC, int sdc, double *pD, int sdd, double *inv_diag_D, int *ipiv)
{
if(m<=0)
return;
const int ps = 4;
int ii, jj, i0, i1, j0, ll, p;
double d1 = 1.0;
double dm1 = -1.0;
// needs to perform row-excanges on the yet-to-be-factorized matrix too
if(pC!=pD)
dgecp_lib(m, n, 1.0, 0, pC, sdc, 0, pD, sdd);
// minimum matrix size
p = n<m ? n : m; // XXX
// main loop
#if defined(TARGET_X64_INTEL_HASWELL)
// 12 columns at a time
jj = 0;
for(; jj<p-11; jj+=12)
{
// pivot & factorize & solve lower
// left block-column
ii = jj;
i0 = ii;
for( ; ii<m-11; ii+=12)
{
kernel_dgemm_nn_12x4_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd);
}
if(m-ii>0)
{
if(m-ii>8)
{
kernel_dgemm_nn_12x4_vs_lib4(jj, &dm1, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd, m-ii, 4);
}
else if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
else
{
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
}
kernel_dgetrf_pivot_4_lib4(m-i0, &pD[jj*ps+i0*sdd], sdd, &inv_diag_D[jj], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+4)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+4)*ps);
}
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+4)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+4)*ps);
}
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+4)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+4)*ps);
}
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+4)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+4)*ps);
}
// middle block-column
ii = i0;
kernel_dtrsm_nn_ll_one_4x4_lib4(ii, &pD[ii*sdd], &pD[(jj+4)*ps], sdd, &pD[(jj+4)*ps+ii*sdd], &pD[(jj+4)*ps+ii*sdd], &pD[ii*ps+ii*sdd]);
ii += 4;
i1 = ii;
for( ; ii<m-11; ii+=12)
{
kernel_dgemm_nn_12x4_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+4)*ps], sdd, &d1, &pD[(jj+4)*ps+ii*sdd], sdd, &pD[(jj+4)*ps+ii*sdd], sdd);
}
if(m-ii>0)
{
if(m-ii>8)
{
kernel_dgemm_nn_12x4_vs_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, &pD[(jj+4)*ps], sdd, &d1, &pD[(jj+4)*ps+ii*sdd], sdd, &pD[(jj+4)*ps+ii*sdd], sdd, m-ii, 4);
}
else if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
else
{
kernel_dgemm_nn_4x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
}
kernel_dgetrf_pivot_4_lib4(m-i1, &pD[(jj+4)*ps+i1*sdd], sdd, &inv_diag_D[(jj+4)], &ipiv[i1]);
ipiv[i1+0] += i1;
if(ipiv[i1+0]!=i1+0)
{
drowsw_lib(jj+4, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps);
drowsw_lib(n-jj-8, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps+(jj+8)*ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps+(jj+8)*ps);
}
ipiv[i1+1] += i1;
if(ipiv[i1+1]!=i1+1)
{
drowsw_lib(jj+4, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps);
drowsw_lib(n-jj-8, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps+(jj+8)*ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps+(jj+8)*ps);
}
ipiv[i1+2] += i1;
if(ipiv[i1+2]!=i1+2)
{
drowsw_lib(jj+4, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps);
drowsw_lib(n-jj-8, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps+(jj+8)*ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps+(jj+8)*ps);
}
ipiv[i1+3] += i1;
if(ipiv[i1+3]!=i1+3)
{
drowsw_lib(jj+4, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps);
drowsw_lib(n-jj-8, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps+(jj+8)*ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps+(jj+8)*ps);
}
// right block-column
ii = i0;
kernel_dtrsm_nn_ll_one_8x4_lib4(ii, &pD[ii*sdd], sdd, &pD[(jj+8)*ps], sdd, &pD[(jj+8)*ps+ii*sdd], sdd, &pD[(jj+8)*ps+ii*sdd], sdd, &pD[ii*ps+ii*sdd], sdd);
ii += 8;
i1 = ii;
for( ; ii<m-11; ii+=12)
{
kernel_dgemm_nn_12x4_lib4((jj+8), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+8)*ps], sdd, &d1, &pD[(jj+8)*ps+ii*sdd], sdd, &pD[(jj+8)*ps+ii*sdd], sdd);
}
if(m-ii>0)
{
if(m-ii>8)
{
kernel_dgemm_nn_12x4_vs_lib4((jj+8), &dm1, &pD[ii*sdd], sdd, &pD[(jj+8)*ps], sdd, &d1, &pD[(jj+8)*ps+ii*sdd], sdd, &pD[(jj+8)*ps+ii*sdd], sdd, m-ii, 4);
}
else if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4((jj+8), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+8)*ps], sdd, &d1, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
else
{
kernel_dgemm_nn_4x4_gen_lib4((jj+8), &dm1, &pD[ii*sdd], 0, &pD[(jj+8)*ps], sdd, &d1, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
}
kernel_dgetrf_pivot_4_lib4(m-i1, &pD[(jj+8)*ps+i1*sdd], sdd, &inv_diag_D[(jj+8)], &ipiv[i1]);
ipiv[i1+0] += i1;
if(ipiv[i1+0]!=i1+0)
{
drowsw_lib(jj+8, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps);
drowsw_lib(n-jj-12, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps+(jj+12)*ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps+(jj+12)*ps);
}
ipiv[i1+1] += i1;
if(ipiv[i1+1]!=i1+1)
{
drowsw_lib(jj+8, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps);
drowsw_lib(n-jj-12, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps+(jj+12)*ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps+(jj+12)*ps);
}
ipiv[i1+2] += i1;
if(ipiv[i1+2]!=i1+2)
{
drowsw_lib(jj+8, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps);
drowsw_lib(n-jj-12, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps+(jj+12)*ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps+(jj+12)*ps);
}
ipiv[i1+3] += i1;
if(ipiv[i1+3]!=i1+3)
{
drowsw_lib(jj+8, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps);
drowsw_lib(n-jj-12, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps+(jj+12)*ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps+(jj+12)*ps);
}
// solve upper
// i0 -= 8; // 4 ???
ll = jj+12;
for( ; ll<n-3; ll+=4)
{
kernel_dtrsm_nn_ll_one_12x4_lib4(i0, &pD[i0*sdd], sdd, &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[i0*ps+i0*sdd], sdd);
}
if(ll<n)
{
kernel_dtrsm_nn_ll_one_12x4_vs_lib4(i0, &pD[i0*sdd], sdd, &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[i0*ps+i0*sdd], sdd, 12, n-ll);
}
}
if(m>=n)
{
if(n-jj>0)
{
if(n-jj<=4)
goto left_n_4;
else if(n-jj<=8)
goto left_n_8;
else
goto left_n_12;
}
}
else // n>m
{
if(m-jj>0)
{
if(m-jj<=4)
goto left_m_4;
else if(m-jj<=8)
goto left_m_8;
else
goto left_m_12;
}
}
#elif defined(TARGET_X64_INTEL_SANDY_BRIDGE)
// 8 columns at a time
jj = 0;
for(; jj<p-7; jj+=8)
{
// pivot & factorize & solve lower
// left block-column
ii = jj;
i0 = ii;
#if defined(TARGET_X64_INTEL_HASWELL) // XXX
for( ; ii<m-11; ii+=12)
{
kernel_dgemm_nn_12x4_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd);
}
if(m-ii>0)
{
if(m-ii>8)
{
kernel_dgemm_nn_12x4_vs_lib4(jj, &dm1, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd, m-ii, 4);
}
else if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
else
{
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
}
#else // SANDY_BRIDGE
for( ; ii<m-7; ii+=8)
{
kernel_dgemm_nn_8x4_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd);
}
if(m-ii>0)
{
if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
else
{
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
}
#endif
kernel_dgetrf_pivot_4_lib4(m-i0, &pD[jj*ps+i0*sdd], sdd, &inv_diag_D[jj], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+4)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+4)*ps);
}
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+4)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+4)*ps);
}
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+4)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+4)*ps);
}
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+4)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+4)*ps);
}
// right block-column
ii = i0;
kernel_dtrsm_nn_ll_one_4x4_lib4(ii, &pD[ii*sdd], &pD[(jj+4)*ps], sdd, &pD[(jj+4)*ps+ii*sdd], &pD[(jj+4)*ps+ii*sdd], &pD[ii*ps+ii*sdd]);
ii += 4;
i0 = ii;
#if defined(TARGET_X64_INTEL_HASWELL) // XXX
for( ; ii<m-11; ii+=12)
{
kernel_dgemm_nn_12x4_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+4)*ps], sdd, &d1, &pD[(jj+4)*ps+ii*sdd], sdd, &pD[(jj+4)*ps+ii*sdd], sdd);
}
if(m-ii>0)
{
if(m-ii>8)
{
kernel_dgemm_nn_12x4_vs_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, &pD[(jj+4)*ps], sdd, &d1, &pD[(jj+4)*ps+ii*sdd], sdd, &pD[(jj+4)*ps+ii*sdd], sdd, m-ii, 4);
}
else if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
else
{
kernel_dgemm_nn_4x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
}
#else // SANDY_BRIDGE
for( ; ii<m-7; ii+=8)
{
kernel_dgemm_nn_8x4_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+4)*ps], sdd, &d1, &pD[(jj+4)*ps+ii*sdd], sdd, &pD[(jj+4)*ps+ii*sdd], sdd);
}
if(m-ii>0)
{
if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
else
{
kernel_dgemm_nn_4x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
}
#endif
kernel_dgetrf_pivot_4_lib4(m-i0, &pD[(jj+4)*ps+i0*sdd], sdd, &inv_diag_D[(jj+4)], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj+4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-8, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+8)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+8)*ps);
}
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj+4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-8, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+8)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+8)*ps);
}
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj+4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-8, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+8)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+8)*ps);
}
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj+4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-8, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+8)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+8)*ps);
}
// solve upper
i0 -= 4;
ll = jj+8;
for( ; ll<n-3; ll+=4)
{
kernel_dtrsm_nn_ll_one_8x4_lib4(i0, &pD[i0*sdd], sdd, &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[i0*ps+i0*sdd], sdd);
}
if(ll<n)
{
kernel_dtrsm_nn_ll_one_8x4_vs_lib4(i0, &pD[i0*sdd], sdd, &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[i0*ps+i0*sdd], sdd, 8, n-ll);
}
}
if(m>=n)
{
if(n-jj>0)
{
if(n-jj<=4) // (m>=1 && n==1) || (m>=2 && n==2) || m>=3 && n==3
{
goto left_n_4;
}
else // (m>=5 && n==5) || (m>=6 && n==6) || (m>=7 && n==7)
goto left_n_8;
}
}
else // n>m
{
if(m-jj>0)
{
if(m-jj<=4) // (m==1 && n>=2) || (m==2 && n>=3) || (m==3 && n>=4) || (m==4 && n>=5)
goto left_m_4;
else // (m==5 && n>=6) || (m==6 && n>=7) || (m==7 && n>=8)
{
goto left_m_8;
}
}
}
#else
// 4 columns at a time
jj = 0;
for(; jj<p-3; jj+=4) // XXX
{
// pivot & factorize & solve lower
ii = jj;
i0 = ii;
#if defined(TARGET_X64_INTEL_HASWELL) // XXX
for( ; ii<m-11; ii+=12)
{
kernel_dgemm_nn_12x4_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd);
}
if(m-ii>0)
{
if(m-ii>8)
{
kernel_dgemm_nn_12x4_vs_lib4(jj, &dm1, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd, m-ii, 4);
}
else if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
else
{
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
}
#elif defined(TARGET_X64_INTEL_SANDY_BRIDGE) // XXX
for( ; ii<m-7; ii+=8)
{
kernel_dgemm_nn_8x4_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd);
}
if(m-ii>0)
{
if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
else
{
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
}
#else
for( ; ii<m-3; ii+=4)
{
kernel_dgemm_nn_4x4_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], &pD[jj*ps+ii*sdd]);
}
if(m-ii>0)
{
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
#endif
kernel_dgetrf_pivot_4_lib4(m-i0, &pD[jj*ps+i0*sdd], sdd, &inv_diag_D[jj], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+4)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+4)*ps);
}
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+4)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+4)*ps);
}
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+4)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+4)*ps);
}
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+4)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+4)*ps);
}
// solve upper
ll = jj+4;
for( ; ll<n-3; ll+=4)
{
kernel_dtrsm_nn_ll_one_4x4_lib4(i0, &pD[i0*sdd], &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], &pD[ll*ps+i0*sdd], &pD[i0*ps+i0*sdd]);
}
if(n-ll>0)
{
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(i0, &pD[i0*sdd], &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], &pD[ll*ps+i0*sdd], &pD[i0*ps+i0*sdd], 4, n-ll);
}
}
if(m>=n)
{
if(n-jj>0)
{
goto left_n_4;
}
}
else
{
if(m-jj>0)
{
goto left_m_4;
}
}
#endif
// common return if jj==n
return;
// clean up
#if defined(TARGET_X64_INTEL_HASWELL)
left_n_12:
// 9-12 columns at a time
// pivot & factorize & solve lower
// left block-column
ii = jj;
i0 = ii;
for( ; ii<m-8; ii+=12)
{
kernel_dgemm_nn_12x4_vs_lib4(jj, &dm1, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd, m-ii, 4);
}
if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
// ii+=8;
}
else if(m-ii>0)
{
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
// ii+=4;
}
kernel_dgetrf_pivot_4_lib4(m-i0, &pD[jj*ps+i0*sdd], sdd, &inv_diag_D[jj], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+4)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+4)*ps);
}
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+4)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+4)*ps);
}
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+4)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+4)*ps);
}
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+4)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+4)*ps);
}
// middle block-column
ii = i0;
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(ii, &pD[ii*sdd], &pD[(jj+4)*ps], sdd, &pD[(jj+4)*ps+ii*sdd], &pD[(jj+4)*ps+ii*sdd], &pD[ii*ps+ii*sdd], 4, n-jj-4);
ii += 4;
i1 = ii;
for( ; ii<m-8; ii+=12)
{
kernel_dgemm_nn_12x4_vs_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, &pD[(jj+4)*ps], sdd, &d1, &pD[(jj+4)*ps+ii*sdd], sdd, &pD[(jj+4)*ps+ii*sdd], sdd, m-ii, n-jj-4);
}
if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj-4);
}
else if(m-ii>0)
{
kernel_dgemm_nn_4x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj-4);
}
kernel_dgetrf_pivot_4_vs_lib4(m-i1, n-jj-4, &pD[(jj+4)*ps+i1*sdd], sdd, &inv_diag_D[(jj+4)], &ipiv[i1]);
ipiv[i1+0] += i1;
if(ipiv[i1+0]!=i1+0)
{
drowsw_lib(jj+4, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps);
drowsw_lib(n-jj-8, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps+(jj+8)*ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps+(jj+8)*ps);
}
if(n-jj-4>1)
{
ipiv[i1+1] += i1;
if(ipiv[i1+1]!=i1+1)
{
drowsw_lib(jj+4, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps);
drowsw_lib(n-jj-8, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps+(jj+8)*ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps+(jj+8)*ps);
}
if(n-jj-4>2)
{
ipiv[i1+2] += i1;
if(ipiv[i1+2]!=i1+2)
{
drowsw_lib(jj+4, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps);
drowsw_lib(n-jj-8, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps+(jj+8)*ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps+(jj+8)*ps);
}
if(n-jj-4>3)
{
ipiv[i1+3] += i1;
if(ipiv[i1+3]!=i1+3)
{
drowsw_lib(jj+4, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps);
drowsw_lib(n-jj-8, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps+(jj+8)*ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps+(jj+8)*ps);
}
}
}
}
// right block-column
ii = i0;
kernel_dtrsm_nn_ll_one_8x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[(jj+8)*ps], sdd, &pD[(jj+8)*ps+ii*sdd], sdd, &pD[(jj+8)*ps+ii*sdd], sdd, &pD[ii*ps+ii*sdd], sdd, 8, n-jj-8);
ii += 8;
i1 = ii;
for( ; ii<m-8; ii+=12)
{
kernel_dgemm_nn_12x4_vs_lib4((jj+8), &dm1, &pD[ii*sdd], sdd, &pD[(jj+8)*ps], sdd, &d1, &pD[(jj+8)*ps+ii*sdd], sdd, &pD[(jj+8)*ps+ii*sdd], sdd, m-ii, n-jj-8);
}
if(m-ii>4)
{
kernel_dgemm_nn_8x4_gen_lib4((jj+8), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+8)*ps], sdd, &d1, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj-8);
}
else if(m-ii>0)
{
kernel_dgemm_nn_4x4_gen_lib4((jj+8), &dm1, &pD[ii*sdd], 0, &pD[(jj+8)*ps], sdd, &d1, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj-8);
}
kernel_dgetrf_pivot_4_vs_lib4(m-i1, n-jj-8, &pD[(jj+8)*ps+i1*sdd], sdd, &inv_diag_D[(jj+8)], &ipiv[i1]);
ipiv[i1+0] += i1;
if(ipiv[i1+0]!=i1+0)
{
drowsw_lib(jj+8, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps);
drowsw_lib(n-jj-12, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps+(jj+12)*ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps+(jj+12)*ps);
}
if(n-jj-8>1)
{
ipiv[i1+1] += i1;
if(ipiv[i1+1]!=i1+1)
{
drowsw_lib(jj+8, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps);
drowsw_lib(n-jj-12, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps+(jj+12)*ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps+(jj+12)*ps);
}
if(n-jj-8>2)
{
ipiv[i1+2] += i1;
if(ipiv[i1+2]!=i1+2)
{
drowsw_lib(jj+8, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps);
drowsw_lib(n-jj-12, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps+(jj+12)*ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps+(jj+12)*ps);
}
if(n-jj-8>3)
{
ipiv[i1+3] += i1;
if(ipiv[i1+3]!=i1+3)
{
drowsw_lib(jj+8, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps);
drowsw_lib(n-jj-12, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps+(jj+12)*ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps+(jj+12)*ps);
}
}
}
}
// solve upper
// there is no upper
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL)
left_m_12:
// 9-12 rows at a time
// pivot & factorize & solve lower
// left block-column
ii = jj;
i0 = ii;
kernel_dgemm_nn_12x4_vs_lib4(jj, &dm1, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, &d1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd, m-ii, 4);
kernel_dgetrf_pivot_4_lib4(m-i0, &pD[jj*ps+i0*sdd], sdd, &inv_diag_D[jj], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+4)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+4)*ps);
}
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+4)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+4)*ps);
}
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+4)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+4)*ps);
}
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+4)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+4)*ps);
}
// middle block-column
ii = i0;
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(ii, &pD[ii*sdd], &pD[(jj+4)*ps], sdd, &pD[(jj+4)*ps+ii*sdd], &pD[(jj+4)*ps+ii*sdd], &pD[ii*ps+ii*sdd], 4, n-jj-4);
ii += 4;
i1 = ii;
kernel_dgemm_nn_8x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj-4);
kernel_dgetrf_pivot_4_vs_lib4(m-i1, n-jj-4, &pD[(jj+4)*ps+i1*sdd], sdd, &inv_diag_D[(jj+4)], &ipiv[i1]);
ipiv[i1+0] += i1;
if(ipiv[i1+0]!=i1+0)
{
drowsw_lib(jj+4, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps);
drowsw_lib(n-jj-8, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps+(jj+8)*ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps+(jj+8)*ps);
}
if(m-jj-4>1)
{
ipiv[i1+1] += i1;
if(ipiv[i1+1]!=i1+1)
{
drowsw_lib(jj+4, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps);
drowsw_lib(n-jj-8, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps+(jj+8)*ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps+(jj+8)*ps);
}
if(m-jj-4>2)
{
ipiv[i1+2] += i1;
if(ipiv[i1+2]!=i1+2)
{
drowsw_lib(jj+4, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps);
drowsw_lib(n-jj-8, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps+(jj+8)*ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps+(jj+8)*ps);
}
if(m-jj-4>3)
{
ipiv[i1+3] += i1;
if(ipiv[i1+3]!=i1+3)
{
drowsw_lib(jj+4, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps);
drowsw_lib(n-jj-8, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps+(jj+8)*ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps+(jj+8)*ps);
}
}
}
}
// right block-column
ii = i0;
kernel_dtrsm_nn_ll_one_8x4_vs_lib4(ii, &pD[ii*sdd], sdd, &pD[(jj+8)*ps], sdd, &pD[(jj+8)*ps+ii*sdd], sdd, &pD[(jj+8)*ps+ii*sdd], sdd, &pD[ii*ps+ii*sdd], sdd, 8, n-jj-8);
ii += 8;
i1 = ii;
kernel_dgemm_nn_4x4_gen_lib4((jj+8), &dm1, &pD[ii*sdd], 0, &pD[(jj+8)*ps], sdd, &d1, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, &pD[(jj+8)*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj-8);
kernel_dgetrf_pivot_4_vs_lib4(m-i1, n-jj-8, &pD[(jj+8)*ps+i1*sdd], sdd, &inv_diag_D[(jj+8)], &ipiv[i1]);
ipiv[i1+0] += i1;
if(ipiv[i1+0]!=i1+0)
{
drowsw_lib(jj+8, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps);
drowsw_lib(n-jj-12, pD+(i1+0)/ps*ps*sdd+(i1+0)%ps+(jj+12)*ps, pD+(ipiv[i1+0])/ps*ps*sdd+(ipiv[i1+0])%ps+(jj+12)*ps);
}
if(m-jj-8>1)
{
ipiv[i1+1] += i1;
if(ipiv[i1+1]!=i1+1)
{
drowsw_lib(jj+8, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps);
drowsw_lib(n-jj-12, pD+(i1+1)/ps*ps*sdd+(i1+1)%ps+(jj+12)*ps, pD+(ipiv[i1+1])/ps*ps*sdd+(ipiv[i1+1])%ps+(jj+12)*ps);
}
if(m-jj-8>2)
{
ipiv[i1+2] += i1;
if(ipiv[i1+2]!=i1+2)
{
drowsw_lib(jj+8, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps);
drowsw_lib(n-jj-12, pD+(i1+2)/ps*ps*sdd+(i1+2)%ps+(jj+12)*ps, pD+(ipiv[i1+2])/ps*ps*sdd+(ipiv[i1+2])%ps+(jj+12)*ps);
}
if(m-jj-8>3)
{
ipiv[i1+3] += i1;
if(ipiv[i1+3]!=i1+3)
{
drowsw_lib(jj+8, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps);
drowsw_lib(n-jj-12, pD+(i1+3)/ps*ps*sdd+(i1+3)%ps+(jj+12)*ps, pD+(ipiv[i1+3])/ps*ps*sdd+(ipiv[i1+3])%ps+(jj+12)*ps);
}
}
}
}
// solve upper
// i0 -= 8;
ll = jj+12;
for( ; ll<n; ll+=4)
{
kernel_dtrsm_nn_ll_one_12x4_vs_lib4(i0, &pD[i0*sdd], sdd, &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[i0*ps+i0*sdd], sdd, m-i0, n-ll);
}
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL) || defined(TARGET_X64_INTEL_SANDY_BRIDGE)
left_n_8:
// 5-8 columns at a time
// pivot & factorize & solve lower
// left block-column
ii = jj;
i0 = ii;
for( ; ii<m-4; ii+=8)
{
kernel_dgemm_nn_8x4_gen_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
}
if(m-ii>0)
{
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
// ii+=4;
}
kernel_dgetrf_pivot_4_lib4(m-i0, &pD[jj*ps+i0*sdd], sdd, &inv_diag_D[jj], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+4)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+4)*ps);
}
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+4)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+4)*ps);
}
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+4)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+4)*ps);
}
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+4)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+4)*ps);
}
// right block-column
ii = i0;
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(ii, &pD[ii*sdd], &pD[(jj+4)*ps], sdd, &pD[(jj+4)*ps+ii*sdd], &pD[(jj+4)*ps+ii*sdd], &pD[ii*ps+ii*sdd], 4, n-jj-4);
ii += 4;
i0 = ii;
for( ; ii<m-4; ii+=8)
{
kernel_dgemm_nn_8x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], sdd, 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], 0, sdd, m-ii, 0, n-jj-4);
}
if(m-ii>0)
{
kernel_dgemm_nn_4x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj-4);
}
kernel_dgetrf_pivot_4_vs_lib4(m-i0, n-jj-4, &pD[(jj+4)*ps+i0*sdd], sdd, &inv_diag_D[(jj+4)], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj+4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-8, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+8)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+8)*ps);
}
if(n-jj-4>1)
{
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj+4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-8, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+8)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+8)*ps);
}
if(n-jj-4>2)
{
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj+4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-8, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+8)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+8)*ps);
}
if(n-jj-4>3)
{
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj+4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-8, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+8)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+8)*ps);
}
}
}
}
// solve upper
// there is no upper
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL) || defined(TARGET_X64_INTEL_SANDY_BRIDGE)
left_m_8:
// 5-8 rows at a time
// pivot & factorize & solve lower
// left block-column
ii = jj;
i0 = ii;
kernel_dgemm_nn_8x4_gen_lib4(jj, &dm1, &pD[ii*sdd], sdd, 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, 4);
kernel_dgetrf_pivot_4_lib4(m-i0, &pD[jj*ps+i0*sdd], sdd, &inv_diag_D[jj], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+4)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+4)*ps);
}
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+4)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+4)*ps);
}
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+4)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+4)*ps);
}
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+4)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+4)*ps);
}
// right block-column
ii = i0;
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(ii, &pD[ii*sdd], &pD[(jj+4)*ps], sdd, &pD[(jj+4)*ps+ii*sdd], &pD[(jj+4)*ps+ii*sdd], &pD[ii*ps+ii*sdd], 4, n-jj-4);
ii += 4;
i0 = ii;
kernel_dgemm_nn_4x4_gen_lib4((jj+4), &dm1, &pD[ii*sdd], 0, &pD[(jj+4)*ps], sdd, &d1, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, &pD[(jj+4)*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj-4);
kernel_dgetrf_pivot_4_vs_lib4(m-i0, n-jj-4, &pD[(jj+4)*ps+i0*sdd], sdd, &inv_diag_D[(jj+4)], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj+4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-8, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+8)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+8)*ps);
}
if(m-jj-4>1)
{
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj+4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-8, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+8)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+8)*ps);
}
if(m-jj-4>2)
{
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj+4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-8, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+8)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+8)*ps);
}
if(m-jj-4>3)
{
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj+4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-8, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+8)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+8)*ps);
}
}
}
}
// solve upper
i0 -= 4;
ll = jj+8;
for( ; ll<n; ll+=4)
{
kernel_dtrsm_nn_ll_one_8x4_vs_lib4(i0, &pD[i0*sdd], sdd, &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[ll*ps+i0*sdd], sdd, &pD[i0*ps+i0*sdd], sdd, m-i0, n-ll);
}
return;
#endif
left_n_4:
// 1-4 columns at a time
// pivot & factorize & solve lower
ii = jj;
i0 = ii;
#if 0//defined(TARGET_X64_AVX2) || defined(TARGET_X64_AVX)
for( ; ii<m-4; ii+=8)
{
kernel_dgemm_nn_8x4_vs_lib4(m-ii, n-jj, jj, &pD[ii*sdd], sdd, &pD[jj*ps], sdd, -1, &pD[jj*ps+ii*sdd], sdd, &pD[jj*ps+ii*sdd], sdd, 0, 0);
}
if(m-ii>0)
{
kernel_dgemm_nn_4x4_vs_lib4(m-ii, n-jj, jj, &pD[ii*sdd], &pD[jj*ps], sdd, -1, &pD[jj*ps+ii*sdd], &pD[jj*ps+ii*sdd], 0, 0);
// ii+=4;
}
#else
for( ; ii<m; ii+=4)
{
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj);
}
#endif
kernel_dgetrf_pivot_4_vs_lib4(m-i0, n-jj, &pD[jj*ps+i0*sdd], sdd, &inv_diag_D[jj], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+4)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+4)*ps);
}
if(n-jj>1)
{
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+4)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+4)*ps);
}
if(n-jj>2)
{
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+4)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+4)*ps);
}
if(n-jj>3)
{
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+4)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+4)*ps);
}
}
}
}
// solve upper
if(0) // there is no upper
{
ll = jj+4;
for( ; ll<n; ll+=4)
{
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(i0, &pD[i0*sdd], &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], &pD[ll*ps+i0*sdd], &pD[i0*ps+i0*sdd], m-i0, n-ll);
}
}
return;
left_m_4:
// 1-4 rows at a time
// pivot & factorize & solve lower
ii = jj;
i0 = ii;
kernel_dgemm_nn_4x4_gen_lib4(jj, &dm1, &pD[ii*sdd], 0, &pD[jj*ps], sdd, &d1, 0, &pD[jj*ps+ii*sdd], sdd, 0, &pD[jj*ps+ii*sdd], sdd, 0, m-ii, 0, n-jj);
kernel_dgetrf_pivot_4_vs_lib4(m-i0, n-jj, &pD[jj*ps+i0*sdd], sdd, &inv_diag_D[jj], &ipiv[i0]);
ipiv[i0+0] += i0;
if(ipiv[i0+0]!=i0+0)
{
drowsw_lib(jj, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps);
drowsw_lib(n-jj-4, pD+(i0+0)/ps*ps*sdd+(i0+0)%ps+(jj+4)*ps, pD+(ipiv[i0+0])/ps*ps*sdd+(ipiv[i0+0])%ps+(jj+4)*ps);
}
if(m-i0>1)
{
ipiv[i0+1] += i0;
if(ipiv[i0+1]!=i0+1)
{
drowsw_lib(jj, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps);
drowsw_lib(n-jj-4, pD+(i0+1)/ps*ps*sdd+(i0+1)%ps+(jj+4)*ps, pD+(ipiv[i0+1])/ps*ps*sdd+(ipiv[i0+1])%ps+(jj+4)*ps);
}
if(m-i0>2)
{
ipiv[i0+2] += i0;
if(ipiv[i0+2]!=i0+2)
{
drowsw_lib(jj, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps);
drowsw_lib(n-jj-4, pD+(i0+2)/ps*ps*sdd+(i0+2)%ps+(jj+4)*ps, pD+(ipiv[i0+2])/ps*ps*sdd+(ipiv[i0+2])%ps+(jj+4)*ps);
}
if(m-i0>3)
{
ipiv[i0+3] += i0;
if(ipiv[i0+3]!=i0+3)
{
drowsw_lib(jj, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps);
drowsw_lib(n-jj-4, pD+(i0+3)/ps*ps*sdd+(i0+3)%ps+(jj+4)*ps, pD+(ipiv[i0+3])/ps*ps*sdd+(ipiv[i0+3])%ps+(jj+4)*ps);
}
}
}
}
// solve upper
ll = jj+4;
for( ; ll<n; ll+=4)
{
kernel_dtrsm_nn_ll_one_4x4_vs_lib4(i0, &pD[i0*sdd], &pD[ll*ps], sdd, &pD[ll*ps+i0*sdd], &pD[ll*ps+i0*sdd], &pD[i0*ps+i0*sdd], m-i0, n-ll);
}
return;
}
# if 0
void dlauum_dpotrf_blk_nt_l_lib(int m, int n, int nv, int *rv, int *cv, double *pA, int sda, double *pB, int sdb, int alg, double *pC, int sdc, double *pD, int sdd, double *inv_diag_D)
{
if(m<=0 || n<=0)
return;
// TODO remove
int k = cv[nv-1];
const int ps = 4;
int i, j, l;
int ii, iii, jj, kii, kiii, kjj, k0, k1;
i = 0;
ii = 0;
iii = 0;
#if defined(TARGET_X64_INTEL_SANDY_BRIDGE) || defined(TARGET_X64_INTEL_HASWELL)
for(; i<m-7; i+=8)
{
while(ii<nv && rv[ii]<i+8)
ii++;
if(ii<nv)
kii = cv[ii];
else
kii = cv[ii-1];
j = 0;
jj = 0;
for(; j<i && j<n-3; j+=4)
{
while(jj<nv && rv[jj]<j+4)
jj++;
if(jj<nv)
kjj = cv[jj];
else
kjj = cv[jj-1];
k0 = kii<kjj ? kii : kjj;
kernel_dgemm_dtrsm_nt_rl_inv_8x4_lib4(k0, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], alg, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &inv_diag_D[j]);
}
if(j<n)
{
while(jj<nv && rv[jj]<j+4)
jj++;
if(jj<nv)
kjj = cv[jj];
else
kjj = cv[jj-1];
k0 = kii<kjj ? kii : kjj;
if(j<i) // dgemm
{
kernel_dgemm_dtrsm_nt_rl_inv_8x4_vs_lib4(k0, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], alg, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &inv_diag_D[j], 8, n-j);
}
else // dsyrk
{
kernel_dsyrk_dpotrf_nt_l_8x4_vs_lib4(k0, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], alg, &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &inv_diag_D[j], 8, n-j);
if(j<n-4)
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k, &pA[(i+4)*sda], &pB[(j+4)*sdb], j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], alg, &pC[(j+4)*ps+(j+4)*sdc], &pD[(j+4)*ps+(j+4)*sdd], &inv_diag_D[j+4], 4, n-j-4); // TODO
}
}
}
}
if(m>i)
{
if(m-i<=4)
{
goto left_4;
}
else
{
goto left_8;
}
}
#else
for(; i<m-3; i+=4)
{
while(ii<nv && rv[ii]<i+4)
ii++;
if(ii<nv)
kii = cv[ii];
else
kii = cv[ii-1];
j = 0;
jj = 0;
for(; j<i && j<n-3; j+=4)
{
while(jj<nv && rv[jj]<j+4)
jj++;
if(jj<nv)
kjj = cv[jj];
else
kjj = cv[jj-1];
k0 = kii<kjj ? kii : kjj;
kernel_dgemm_dtrsm_nt_rl_inv_4x4_lib4(k0, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], alg, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &inv_diag_D[j]);
}
if(j<n)
{
while(jj<nv && rv[jj]<j+4)
jj++;
if(jj<nv)
kjj = cv[jj];
else
kjj = cv[jj-1];
k0 = kii<kjj ? kii : kjj;
if(i<j) // dgemm
{
kernel_dgemm_dtrsm_nt_rl_inv_4x4_vs_lib4(k0, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], alg, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &inv_diag_D[j], 4, n-j);
}
else // dsyrk
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k0, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], alg, &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &inv_diag_D[j], 4, n-j);
}
}
}
if(m>i)
{
goto left_4;
}
#endif
// common return if i==m
return;
// clean up loops definitions
#if defined(TARGET_X64_INTEL_SANDY_BRIDGE) || defined(TARGET_X64_INTEL_HASWELL)
left_8:
kii = cv[nv-1];
j = 0;
jj = 0;
for(; j<i && j<n-3; j+=4)
{
while(jj<nv && rv[jj]<j+4)
jj++;
if(jj<nv)
kjj = cv[jj];
else
kjj = cv[jj-1];
k0 = kii<kjj ? kii : kjj;
kernel_dgemm_dtrsm_nt_rl_inv_8x4_vs_lib4(k0, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], alg, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &inv_diag_D[j], m-i, n-j);
}
if(j<n)
{
while(jj<nv && rv[jj]<j+4)
jj++;
if(jj<nv)
kjj = cv[jj];
else
kjj = cv[jj-1];
k0 = kii<kjj ? kii : kjj;
if(j<i) // dgemm
{
kernel_dgemm_dtrsm_nt_rl_inv_8x4_vs_lib4(k0, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], alg, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &inv_diag_D[j], m-i, n-j);
}
else // dsyrk
{
kernel_dsyrk_dpotrf_nt_l_8x4_vs_lib4(k0, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], alg, &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &inv_diag_D[j], m-i, n-j);
if(j<n-4)
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k, &pA[(i+4)*sda], &pB[(j+4)*sdb], j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], alg, &pC[(j+4)*ps+(j+4)*sdc], &pD[(j+4)*ps+(j+4)*sdd], &inv_diag_D[j+4], m-i-4, n-j-4); // TODO
}
}
}
return;
#endif
left_4:
kii = cv[nv-1];
j = 0;
jj = 0;
for(; j<i && j<n-3; j+=4)
{
while(jj<nv && rv[jj]<j+4)
jj++;
if(jj<nv)
kjj = cv[jj];
else
kjj = cv[jj-1];
k0 = kii<kjj ? kii : kjj;
kernel_dgemm_dtrsm_nt_rl_inv_4x4_vs_lib4(k0, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], alg, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &inv_diag_D[j], m-i, n-j);
}
if(j<n)
{
while(jj<nv && rv[jj]<j+4)
jj++;
if(jj<nv)
kjj = cv[jj];
else
kjj = cv[jj-1];
k0 = kii<kjj ? kii : kjj;
if(j<i) // dgemm
{
kernel_dgemm_dtrsm_nt_rl_inv_4x4_vs_lib4(k0, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], alg, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &inv_diag_D[j], m-i, n-j);
}
else // dsyrk
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k0, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], alg, &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &inv_diag_D[j], m-i, n-j);
}
}
return;
}
#endif
/****************************
* new interface
****************************/
#if defined(LA_HIGH_PERFORMANCE)
// dpotrf
void dpotrf_l_libstr(int m, struct d_strmat *sC, int ci, int cj, struct d_strmat *sD, int di, int dj)
{
if(m<=0)
return;
if(ci!=0 | di!=0)
{
printf("\ndpotrf_l_libstr: feature not implemented yet: ci=%d, di=%d\n", ci, di);
exit(1);
}
const int ps = 4;
int sdc = sC->cn;
int sdd = sD->cn;
double *pC = sC->pA + cj*ps;
double *pD = sD->pA + dj*ps;
double *dD = sD->dA;
if(di==0 & dj==0) // XXX what to do if di and dj are not zero
sD->use_dA = 1;
else
sD->use_dA = 0;
int i, j, l;
i = 0;
#if defined(TARGET_X64_INTEL_HASWELL)
for(; i<m-11; i+=12)
{
j = 0;
for(; j<i; j+=4)
{
kernel_dtrsm_nt_rl_inv_12x4_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j]);
}
kernel_dpotrf_nt_l_12x4_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j]);
kernel_dpotrf_nt_l_8x8_lib4(j+4, &pD[(i+4)*sdd], sdd, &pD[(j+4)*sdd], sdd, &pC[(j+4)*ps+(i+4)*sdc], sdc, &pD[(j+4)*ps+(i+4)*sdd], sdd, &dD[j+4]);
}
if(m>i)
{
if(m-i<=4)
{
goto left_4;
}
else if(m-i<=8)
{
goto left_8;
}
else
{
goto left_12;
}
}
#elif defined(TARGET_X64_INTEL_SANDY_BRIDGE)
for(; i<m-7; i+=8)
{
j = 0;
for(; j<i; j+=4)
{
kernel_dtrsm_nt_rl_inv_8x4_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j]);
}
kernel_dpotrf_nt_l_8x4_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j]);
kernel_dpotrf_nt_l_4x4_lib4(j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4]);
}
if(m>i)
{
if(m-i<=4)
{
goto left_4;
}
else
{
goto left_8;
}
}
#else
for(; i<m-3; i+=4)
{
j = 0;
for(; j<i; j+=4)
{
kernel_dtrsm_nt_rl_inv_4x4_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j]);
}
kernel_dpotrf_nt_l_4x4_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j]);
}
if(m>i)
{
goto left_4;
}
#endif
// common return if i==m
return;
// clean up loops definitions
#if defined(TARGET_X64_INTEL_HASWELL)
left_12: // 9 - 12
j = 0;
for(; j<i; j+=4)
{
kernel_dtrsm_nt_rl_inv_12x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, m-j);
}
kernel_dpotrf_nt_l_12x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, m-j);
kernel_dpotrf_nt_l_8x8_vs_lib4(j+4, &pD[(i+4)*sdd], sdd, &pD[(j+4)*sdd], sdd, &pC[(j+4)*ps+(i+4)*sdc], sdc, &pD[(j+4)*ps+(i+4)*sdd], sdd, &dD[j+4], m-i-4, m-j-4);
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL)
left_8:
j = 0;
for(; j<i-8; j+=12)
{
kernel_dtrsm_nt_rl_inv_8x8l_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], sdd, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, m-j);
kernel_dtrsm_nt_rl_inv_8x8u_vs_lib4((j+4), &pD[i*sdd], sdd, &pD[(j+4)*sdd], sdd, &pC[(j+4)*ps+i*sdc], sdc, &pD[(j+4)*ps+i*sdd], sdd, &pD[(j+4)*ps+(j+4)*sdd], sdd, &dD[(j+4)], m-i, m-(j+4));
}
if(j<i-4)
{
kernel_dtrsm_nt_rl_inv_8x8l_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], sdd, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, m-j);
kernel_dtrsm_nt_rl_inv_4x4_vs_lib4((j+4), &pD[i*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+i*sdc], &pD[(j+4)*ps+i*sdd], &pD[(j+4)*ps+(j+4)*sdd], &dD[(j+4)], m-i, m-(j+4));
j += 8;
}
else if(j<i)
{
kernel_dtrsm_nt_rl_inv_8x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, m-j);
j += 4;
}
kernel_dpotrf_nt_l_8x8_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], sdd, &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, m-j);
return;
#endif
#if defined(TARGET_X64_INTEL_SANDY_BRIDGE)
left_8:
j = 0;
for(; j<i; j+=4)
{
kernel_dtrsm_nt_rl_inv_8x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, m-j);
}
kernel_dpotrf_nt_l_8x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, m-j);
kernel_dpotrf_nt_l_4x4_vs_lib4(j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4], m-i-4, m-j-4);
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL)
left_4:
j = 0;
for(; j<i-8; j+=12)
{
kernel_dtrsm_nt_rl_inv_4x12_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], sdd, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], sdd, &dD[j], m-i, m-j);
}
if(j<i-4)
{
kernel_dtrsm_nt_rl_inv_4x8_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], sdd, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], sdd, &dD[j], m-i, m-j);
j += 8;
}
else if(j<i)
{
kernel_dtrsm_nt_rl_inv_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j], m-i, m-j);
j += 4;
}
kernel_dpotrf_nt_l_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j], m-i, m-j);
return;
#else
left_4:
j = 0;
for(; j<i; j+=4)
{
kernel_dtrsm_nt_rl_inv_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j], m-i, m-j);
}
kernel_dpotrf_nt_l_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j], m-i, m-j);
return;
#endif
}
// dpotrf
void dpotrf_l_mn_libstr(int m, int n, struct d_strmat *sC, int ci, int cj, struct d_strmat *sD, int di, int dj)
{
if(m<=0 || n<=0)
return;
if(ci!=0 | di!=0)
{
printf("\ndpotrf_l_mn_libstr: feature not implemented yet: ci=%d, di=%d\n", ci, di);
exit(1);
}
const int ps = 4;
int sdc = sC->cn;
int sdd = sD->cn;
double *pC = sC->pA + cj*ps;
double *pD = sD->pA + dj*ps;
double *dD = sD->dA;
if(di==0 & dj==0) // XXX what to do if di and dj are not zero
sD->use_dA = 1;
else
sD->use_dA = 0;
int i, j, l;
i = 0;
#if defined(TARGET_X64_INTEL_HASWELL)
for(; i<m-11; i+=12)
{
j = 0;
for(; j<i & j<n-3; j+=4)
{
kernel_dtrsm_nt_rl_inv_12x4_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j]);
}
if(j<n)
{
if(j<i) // dtrsm
{
kernel_dtrsm_nt_rl_inv_12x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
else // dpptrf
{
if(n<j-11)
{
kernel_dpotrf_nt_l_12x4_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j]);
kernel_dpotrf_nt_l_8x8_lib4(j+4, &pD[(i+4)*sdd], sdd, &pD[(j+4)*sdd], sdd, &pC[(j+4)*ps+(i+4)*sdc], sdc, &pD[(j+4)*ps+(i+4)*sdd], sdd, &dD[j+4]);
}
else
{
kernel_dpotrf_nt_l_12x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
kernel_dpotrf_nt_l_8x4_vs_lib4(j+4, &pD[(i+4)*sdd], sdd, &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], sdc, &pD[(j+4)*ps+(i+4)*sdd], sdd, &dD[j+4], m-i-4, n-j-4);
if(j<n-8)
{
kernel_dpotrf_nt_l_4x4_vs_lib4(j+8, &pD[(i+8)*sdd], &pD[(j+8)*sdd], &pC[(j+8)*ps+(i+8)*sdc], &pD[(j+8)*ps+(i+8)*sdd], &dD[j+8], m-i-8, n-j-8);
}
}
}
}
}
}
if(m>i)
{
if(m-i<=4)
{
goto left_4;
}
else if(m-i<=8)
{
goto left_8;
}
else
{
goto left_12;
}
}
#elif defined(TARGET_X64_INTEL_SANDY_BRIDGE)
for(; i<m-7; i+=8)
{
j = 0;
for(; j<i & j<n-3; j+=4)
{
kernel_dtrsm_nt_rl_inv_8x4_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j]);
}
if(j<n)
{
if(j<i) // dtrsm
{
kernel_dtrsm_nt_rl_inv_8x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
else // dpotrf
{
if(j<n-7)
// if(0)
{
kernel_dpotrf_nt_l_8x4_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j]);
kernel_dpotrf_nt_l_4x4_lib4(j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4]);
}
else
{
kernel_dpotrf_nt_l_8x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
kernel_dpotrf_nt_l_4x4_vs_lib4(j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4], m-i-4, n-j-4);
}
}
}
}
}
if(m>i)
{
if(m-i<=4)
{
goto left_4;
}
else
{
goto left_8;
}
}
#else
for(; i<m-3; i+=4)
{
j = 0;
for(; j<i & j<n-3; j+=4)
{
kernel_dtrsm_nt_rl_inv_4x4_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j]);
}
if(j<n)
{
if(i<j) // dtrsm
{
kernel_dtrsm_nt_rl_inv_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
else // dpotrf
{
if(j<n-3)
{
kernel_dpotrf_nt_l_4x4_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j]);
}
else
{
kernel_dpotrf_nt_l_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
}
}
}
if(m>i)
{
goto left_4;
}
#endif
// common return if i==m
return;
// clean up loops definitions
#if defined(TARGET_X64_INTEL_HASWELL)
left_12:
j = 0;
for(; j<i & j<n; j+=4)
{
kernel_dtrsm_nt_rl_inv_12x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
if(j<n)
{
kernel_dpotrf_nt_l_12x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
kernel_dpotrf_nt_l_8x4_vs_lib4(j+4, &pD[(i+4)*sdd], sdd, &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], sdc, &pD[(j+4)*ps+(i+4)*sdd], sdd, &dD[j+4], m-i-4, n-j-4);
if(j<n-8)
{
kernel_dpotrf_nt_l_4x4_vs_lib4(j+8, &pD[(i+8)*sdd], &pD[(j+8)*sdd], &pC[(j+8)*ps+(i+8)*sdc], &pD[(j+8)*ps+(i+8)*sdd], &dD[j+8], m-i-8, n-j-8);
}
}
}
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL)
left_8:
j = 0;
for(; j<i-8 & j<n-8; j+=12)
{
kernel_dtrsm_nt_rl_inv_8x8l_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], sdd, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
kernel_dtrsm_nt_rl_inv_8x8u_vs_lib4((j+4), &pD[i*sdd], sdd, &pD[(j+4)*sdd], sdd, &pC[(j+4)*ps+i*sdc], sdc, &pD[(j+4)*ps+i*sdd], sdd, &pD[(j+4)*ps+(j+4)*sdd], sdd, &dD[(j+4)], m-i, n-(j+4));
}
if(j<i-4 & j<n-4)
{
kernel_dtrsm_nt_rl_inv_8x8l_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], sdd, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
kernel_dtrsm_nt_rl_inv_4x4_vs_lib4((j+4), &pD[i*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+i*sdc], &pD[(j+4)*ps+i*sdd], &pD[(j+4)*ps+(j+4)*sdd], &dD[(j+4)], m-i, n-(j+4));
j += 8;
}
else if(j<i & j<n)
{
kernel_dtrsm_nt_rl_inv_8x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
j += 4;
}
if(j<n)
{
kernel_dpotrf_nt_l_8x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
kernel_dpotrf_nt_l_4x4_vs_lib4(j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4], m-i-4, n-j-4);
}
}
return;
#elif defined(TARGET_X64_INTEL_SANDY_BRIDGE)
left_8:
j = 0;
for(; j<i & j<n; j+=4)
{
kernel_dtrsm_nt_rl_inv_8x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
if(j<n)
{
kernel_dpotrf_nt_l_8x4_vs_lib4(j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
kernel_dpotrf_nt_l_4x4_vs_lib4(j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4], m-i-4, n-j-4);
}
}
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL)
left_4:
j = 0;
for(; j<i-8 & j<n-8; j+=12)
{
kernel_dtrsm_nt_rl_inv_4x12_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], sdd, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
}
if(j<i-4 & j<n-4)
{
kernel_dtrsm_nt_rl_inv_4x8_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], sdd, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
j += 8;
}
else if(j<i & j<n)
{
kernel_dtrsm_nt_rl_inv_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
j += 4;
}
if(j<n)
{
kernel_dpotrf_nt_l_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
return;
#else
left_4:
j = 0;
for(; j<i & j<n; j+=4)
{
kernel_dtrsm_nt_rl_inv_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
if(j<n)
{
kernel_dpotrf_nt_l_4x4_vs_lib4(j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
return;
#endif
}
// dsyrk dpotrf
void dsyrk_dpotrf_ln_libstr(int m, int n, int k, struct d_strmat *sA, int ai, int aj, struct d_strmat *sB, int bi, int bj, struct d_strmat *sC, int ci, int cj, struct d_strmat *sD, int di, int dj)
{
if(m<=0 || n<=0)
return;
if(ai!=0 | bi!=0 | ci!=0 | di!=0)
{
printf("\ndsyrk_dpotrf_ln_libstr: feature not implemented yet: ai=%d, bi=%d, ci=%d, di=%d\n", ai, bi, ci, di);
exit(1);
}
const int ps = 4;
int sda = sA->cn;
int sdb = sB->cn;
int sdc = sC->cn;
int sdd = sD->cn;
double *pA = sA->pA + aj*ps;
double *pB = sB->pA + bj*ps;
double *pC = sC->pA + cj*ps;
double *pD = sD->pA + dj*ps;
double *dD = sD->dA; // XXX what to do if di and dj are not zero
if(di==0 & dj==0)
sD->use_dA = 1;
else
sD->use_dA = 0;
int i, j, l;
i = 0;
#if defined(TARGET_X64_INTEL_HASWELL)
for(; i<m-11; i+=12)
{
j = 0;
for(; j<i & j<n-3; j+=4)
{
kernel_dgemm_dtrsm_nt_rl_inv_12x4_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j]);
}
if(j<n)
{
if(j<i) // dgemm
{
kernel_dgemm_dtrsm_nt_rl_inv_12x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
else // dsyrk
{
if(j<n-11)
{
kernel_dsyrk_dpotrf_nt_l_12x4_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j]);
kernel_dsyrk_dpotrf_nt_l_8x8_lib4(k, &pA[(i+4)*sda], sda, &pB[(j+4)*sdb], sdb, j+4, &pD[(i+4)*sdd], sdd, &pD[(j+4)*sdd], sdd, &pC[(j+4)*ps+(i+4)*sdc], sdc, &pD[(j+4)*ps+(i+4)*sdd], sdd, &dD[j+4]);
}
else
{
kernel_dsyrk_dpotrf_nt_l_12x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
if(j<n-8)
{
kernel_dsyrk_dpotrf_nt_l_8x8_vs_lib4(k, &pA[(i+4)*sda], sda, &pB[(j+4)*sdb], sdb, j+4, &pD[(i+4)*sdd], sdd, &pD[(j+4)*sdd], sdd, &pC[(j+4)*ps+(i+4)*sdc], sdc, &pD[(j+4)*ps+(i+4)*sdd], sdd, &dD[j+4], m-i-4, n-j-4);
}
else
{
kernel_dsyrk_dpotrf_nt_l_8x4_vs_lib4(k, &pA[(i+4)*sda], sda, &pB[(j+4)*sdb], j+4, &pD[(i+4)*sdd], sdd, &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], sdc, &pD[(j+4)*ps+(i+4)*sdd], sdd, &dD[j+4], m-i-4, n-j-4);
}
}
}
}
}
}
if(m>i)
{
if(m-i<=4)
{
goto left_4;
}
else if(m-i<=8)
{
goto left_8;
}
else
{
goto left_12;
}
}
#elif defined(TARGET_X64_INTEL_SANDY_BRIDGE)
for(; i<m-7; i+=8)
{
j = 0;
for(; j<i & j<n-3; j+=4)
{
kernel_dgemm_dtrsm_nt_rl_inv_8x4_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j]);
}
if(j<n)
{
if(j<i) // dgemm
{
kernel_dgemm_dtrsm_nt_rl_inv_8x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
else // dsyrk
{
if(j<n-7)
// if(0)
{
kernel_dsyrk_dpotrf_nt_l_8x4_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j]);
kernel_dsyrk_dpotrf_nt_l_4x4_lib4(k, &pA[(i+4)*sda], &pB[(j+4)*sdb], j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4]);
}
else
{
kernel_dsyrk_dpotrf_nt_l_8x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k, &pA[(i+4)*sda], &pB[(j+4)*sdb], j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4], m-i-4, n-j-4);
}
}
}
}
}
if(m>i)
{
if(m-i<=4)
{
goto left_4;
}
else
{
goto left_8;
}
}
#else
for(; i<m-3; i+=4)
{
j = 0;
for(; j<i & j<n-3; j+=4)
{
kernel_dgemm_dtrsm_nt_rl_inv_4x4_lib4(k, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j]);
}
if(j<n)
{
if(i<j) // dgemm
{
kernel_dgemm_dtrsm_nt_rl_inv_4x4_vs_lib4(k, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
else // dsyrk
{
if(j<n-3)
{
kernel_dsyrk_dpotrf_nt_l_4x4_lib4(k, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j]);
}
else
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
}
}
}
if(m>i)
{
goto left_4;
}
#endif
// common return if i==m
return;
// clean up loops definitions
#if defined(TARGET_X64_INTEL_HASWELL)
left_12:
j = 0;
for(; j<i & j<n; j+=4)
{
kernel_dgemm_dtrsm_nt_rl_inv_12x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
if(j<n)
{
kernel_dsyrk_dpotrf_nt_l_12x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
kernel_dsyrk_dpotrf_nt_l_8x4_vs_lib4(k, &pA[(i+4)*sda], sda, &pB[(j+4)*sdb], j+4, &pD[(i+4)*sdd], sdd, &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], sdc, &pD[(j+4)*ps+(i+4)*sdd], sdd, &dD[j+4], m-i-4, n-j-4);
if(j<n-8)
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k, &pA[(i+8)*sda], &pB[(j+8)*sdb], j+8, &pD[(i+8)*sdd], &pD[(j+8)*sdd], &pC[(j+8)*ps+(i+8)*sdc], &pD[(j+8)*ps+(i+8)*sdd], &dD[j+8], m-i-8, n-j-8);
}
}
}
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL)
left_8:
j = 0;
for(; j<i-8 & j<n-8; j+=12)
{
kernel_dgemm_dtrsm_nt_rl_inv_8x8l_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], sdb, j, &pD[i*sdd], sdd, &pD[j*sdd], sdd, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
kernel_dgemm_dtrsm_nt_rl_inv_8x8u_vs_lib4(k, &pA[i*sda], sda, &pB[(j+4)*sdb], sdb, (j+4), &pD[i*sdd], sdd, &pD[(j+4)*sdd], sdd, &pC[(j+4)*ps+i*sdc], sdc, &pD[(j+4)*ps+i*sdd], sdd, &pD[(j+4)*ps+(j+4)*sdd], sdd, &dD[(j+4)], m-i, n-(j+4));
}
if(j<i-3 & j<n-3)
{
kernel_dgemm_dtrsm_nt_rl_inv_8x8l_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], sdb, j, &pD[i*sdd], sdd, &pD[j*sdd], sdd, &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
kernel_dgemm_dtrsm_nt_rl_inv_4x4_vs_lib4(k, &pA[i*sda], &pB[(j+4)*sdb], (j+4), &pD[i*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+i*sdc], &pD[(j+4)*ps+i*sdd], &pD[(j+4)*ps+(j+4)*sdd], &dD[(j+4)], m-i, n-(j+4));
j += 8;
}
else if(j<i & j<n)
{
kernel_dgemm_dtrsm_nt_rl_inv_8x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
j += 4;
}
if(j<n)
{
kernel_dsyrk_dpotrf_nt_l_8x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k, &pA[(i+4)*sda], &pB[(j+4)*sdb], j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4], m-i-4, n-j-4);
}
}
return;
#elif defined(TARGET_X64_INTEL_SANDY_BRIDGE)
left_8:
j = 0;
for(; j<i & j<n; j+=4)
{
kernel_dgemm_dtrsm_nt_rl_inv_8x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+i*sdc], sdc, &pD[j*ps+i*sdd], sdd, &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
if(j<n)
{
kernel_dsyrk_dpotrf_nt_l_8x4_vs_lib4(k, &pA[i*sda], sda, &pB[j*sdb], j, &pD[i*sdd], sdd, &pD[j*sdd], &pC[j*ps+j*sdc], sdc, &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
if(j<n-4)
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k, &pA[(i+4)*sda], &pB[(j+4)*sdb], j+4, &pD[(i+4)*sdd], &pD[(j+4)*sdd], &pC[(j+4)*ps+(i+4)*sdc], &pD[(j+4)*ps+(i+4)*sdd], &dD[j+4], m-i-4, n-j-4);
}
}
return;
#endif
#if defined(TARGET_X64_INTEL_HASWELL)
left_4:
j = 0;
for(; j<i-8 & j<n-8; j+=12)
{
kernel_dgemm_dtrsm_nt_rl_inv_4x12_vs_lib4(k, &pA[i*sda], &pB[j*sdb], sdb, j, &pD[i*sdd], &pD[j*sdd], sdd, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
}
if(j<i-4 & j<n-4)
{
kernel_dgemm_dtrsm_nt_rl_inv_4x8_vs_lib4(k, &pA[i*sda], &pB[j*sdb], sdb, j, &pD[i*sdd], &pD[j*sdd], sdd, &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], sdd, &dD[j], m-i, n-j);
j += 8;
}
else if(j<i & j<n)
{
kernel_dgemm_dtrsm_nt_rl_inv_4x4_vs_lib4(k, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
j += 4;
}
if(j<n)
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
#else
left_4:
j = 0;
for(; j<i & j<n; j+=4)
{
kernel_dgemm_dtrsm_nt_rl_inv_4x4_vs_lib4(k, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+i*sdc], &pD[j*ps+i*sdd], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
if(j<n)
{
kernel_dsyrk_dpotrf_nt_l_4x4_vs_lib4(k, &pA[i*sda], &pB[j*sdb], j, &pD[i*sdd], &pD[j*sdd], &pC[j*ps+j*sdc], &pD[j*ps+j*sdd], &dD[j], m-i, n-j);
}
#endif
return;
}
// dgetrf without pivoting
void dgetrf_nopivot_libstr(int m, int n, struct d_strmat *sC, int ci, int cj, struct d_strmat *sD, int di, int dj)
{
if(ci!=0 | di!=0)
{
printf("\ndgetf_nopivot_libstr: feature not implemented yet: ci=%d, di=%d\n", ci, di);
exit(1);
}
const int ps = 4;
int sdc = sC->cn;
int sdd = sD->cn;
double *pC = sC->pA + cj*ps;
double *pD = sD->pA + dj*ps;
double *dD = sD->dA; // XXX what to do if di and dj are not zero
dgetrf_nn_nopivot_lib(m, n, pC, sdc, pD, sdd, dD);
if(di==0 && dj==0)
sD->use_dA = 1;
else
sD->use_dA = 0;
return;
}
// dgetrf pivoting
void dgetrf_libstr(int m, int n, struct d_strmat *sC, int ci, int cj, struct d_strmat *sD, int di, int dj, int *ipiv)
{
if(ci!=0 | di!=0)
{
printf("\ndgetrf_libstr: feature not implemented yet: ci=%d, di=%d\n", ci, di);
exit(1);
}
const int ps = 4;
int sdc = sC->cn;
int sdd = sD->cn;
double *pC = sC->pA + cj*ps;
double *pD = sD->pA + dj*ps;
double *dD = sD->dA; // XXX what to do if di and dj are not zero
dgetrf_nn_lib(m, n, pC, sdc, pD, sdd, dD, ipiv);
if(di==0 && dj==0)
sD->use_dA = 1;
else
sD->use_dA = 0;
return;
}
int dgeqrf_work_size_libstr(int m, int n)
{
const int ps = 4;
int cm = (m+ps-1)/ps*ps;
int cn = (n+ps-1)/ps*ps;
return ps*(cm+cn)*sizeof(double);
// return 0;
}
void dgeqrf_libstr(int m, int n, struct d_strmat *sC, int ci, int cj, struct d_strmat *sD, int di, int dj, void *v_work)
{
char *work = (char *) v_work;
if(m<=0 | n<=0)
return;
const int ps = 4;
int sdc = sC->cn;
int sdd = sD->cn;
double *pC = &(DMATEL_LIBSTR(sC,ci,cj));
double *pD = &(DMATEL_LIBSTR(sD,di,dj));
double *dD = sD->dA + di;
int cm = (m+ps-1)/ps*ps;
int cn = (n+ps-1)/ps*ps;
double *pVt = (double *) work;
work += ps*cm*sizeof(double);
double *pW = (double *) work;
work += ps*cn*sizeof(double);
if(pC!=pD)
dgecp_lib(m, n, 1.0, ci&(ps-1), pC, sdc, di&(ps-1), pD, sdd);
int ii;
int imax0 = (ps-(di&(ps-1)))&(ps-1);
int imax = m<n ? m : n;
imax0 = imax<imax0 ? imax : imax0;
if(imax0>0)
{
kernel_dgeqrf_vs_lib4(m, n, imax0, di&(ps-1), pD, sdd, dD);
pD += imax0-ps+ps*sdd+imax0*ps;
dD += imax0;
m -= imax0;
n -= imax0;
imax -= imax0;
}
for(ii=0; ii<imax-3; ii+=4)
{
kernel_dgeqrf_4_lib4(m-ii, pD+ii*sdd+ii*ps, sdd, dD+ii);
#if 0
kernel_dlarf_4_lib4(m-ii, n-ii-4, pD+ii*sdd+ii*ps, sdd, dD+ii, pD+ii*sdd+(ii+4)*ps, sdd);
#else
kernel_dgetr_4_0_lib4(m-ii, pD+ii*sdd+ii*ps, sdd, pVt);
pVt[0+ps*0] = 1.0;
pVt[1+ps*0] = 0.0;
pVt[2+ps*0] = 0.0;
pVt[3+ps*0] = 0.0;
pVt[1+ps*1] = 1.0;
pVt[2+ps*1] = 0.0;
pVt[3+ps*1] = 0.0;
pVt[2+ps*2] = 1.0;
pVt[3+ps*2] = 0.0;
pVt[3+ps*3] = 1.0;
kernel_dlarf_t_4_lib4(m-ii, n-ii-4, pD+ii*sdd+ii*ps, sdd, pVt, dD+ii, pD+ii*sdd+(ii+4)*ps, sdd, pW);
#endif
}
if(ii<imax)
{
kernel_dgeqrf_vs_lib4(m-ii, n-ii, imax-ii, ii&(ps-1), pD+ii*sdd+ii*ps, sdd, dD+ii);
}
return;
}
int dgelqf_work_size_libstr(int m, int n)
{
return 0;
}
void dgelqf_libstr(int m, int n, struct d_strmat *sC, int ci, int cj, struct d_strmat *sD, int di, int dj, void *work)
{
if(m<=0 | n<=0)
return;
const int ps = 4;
int sdc = sC->cn;
int sdd = sD->cn;
double *pC = &(DMATEL_LIBSTR(sC,ci,cj));
double *pD = &(DMATEL_LIBSTR(sD,di,dj));
double *dD = sD->dA + di;
#if defined(TARGET_X64_INTEL_HASWELL)
double pT[144] __attribute__ ((aligned (64))) = {0};
double pK[96] __attribute__ ((aligned (64))) = {0};
#else
double pT[144] = {0};
double pK[96] = {0};
#endif
if(pC!=pD)
dgecp_lib(m, n, 1.0, ci&(ps-1), pC, sdc, di&(ps-1), pD, sdd);
int ii, jj, ll;
int imax0 = (ps-(di&(ps-1)))&(ps-1);
int imax = m<n ? m : n;
#if 0
kernel_dgelqf_vs_lib4(m, n, imax, di&(ps-1), pD, sdd, dD);
#else
imax0 = imax<imax0 ? imax : imax0;
if(imax0>0)
{
kernel_dgelqf_vs_lib4(m, n, imax0, di&(ps-1), pD, sdd, dD);
pD += imax0-ps+ps*sdd+imax0*ps;
dD += imax0;
m -= imax0;
n -= imax0;
imax -= imax0;
}
ii = 0;
#if defined(TARGET_X64_INTEL_HASWELL)
// for(; ii<imax-11; ii+=12)
for(; ii<imax-127; ii+=12) // crossover point ~ ii=128
{
kernel_dgelqf_dlarft12_12_lib4(n-(ii+0), pD+(ii+0)*sdd+(ii+0)*ps, sdd, dD+(ii+0), &pT[0+0*12+0*ps]);
jj = ii+12;
for(; jj<m; jj+=4)
{
kernel_dlarfb12_r_4_lib4(n-ii, pD+ii*sdd+ii*ps, sdd, pT, pD+jj*sdd+ii*ps, pK, m-jj);
}
}
for(; ii<imax-11; ii+=4)
{
kernel_dgelqf_dlarft4_12_lib4(n-ii, pD+ii*sdd+ii*ps, sdd, dD+ii, pT);
jj = ii+12;
for(; jj<m-11; jj+=12)
{
kernel_dlarfb4_r_12_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps, sdd);
}
for(; jj<m-7; jj+=8)
{
kernel_dlarfb4_r_8_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps, sdd);
}
for(; jj<m-3; jj+=4)
{
kernel_dlarfb4_r_4_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps);
}
for(ll=0; ll<m-jj; ll++)
{
kernel_dlarfb4_r_1_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+ll+jj*sdd+ii*ps);
}
}
// 8 9 10 11
if(ii<imax-7)
{
kernel_dgelqf_dlarft4_8_lib4(n-ii, pD+ii*sdd+ii*ps, sdd, dD+ii, pT);
jj = ii+8;
if(jj<m)
{
for(; jj<m-11; jj+=12)
{
kernel_dlarfb4_r_12_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps, sdd);
}
for(; jj<m-7; jj+=8)
{
kernel_dlarfb4_r_8_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps, sdd);
}
for(; jj<m-3; jj+=4)
{
kernel_dlarfb4_r_4_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps);
}
for(ll=0; ll<m-jj; ll++)
{
kernel_dlarfb4_r_1_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+ll+jj*sdd+ii*ps);
}
}
ii += 4;
}
// 4 5 6 7
if(ii<imax-3)
{
kernel_dgelqf_dlarft4_4_lib4(n-ii, pD+ii*sdd+ii*ps, dD+ii, pT);
jj = ii+4;
if(jj<m)
{
for(; jj<m-11; jj+=12)
{
kernel_dlarfb4_r_12_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps, sdd);
}
for(; jj<m-7; jj+=8)
{
kernel_dlarfb4_r_8_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps, sdd);
}
for(; jj<m-3; jj+=4)
{
kernel_dlarfb4_r_4_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps);
}
for(ll=0; ll<m-jj; ll++)
{
kernel_dlarfb4_r_1_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+ll+jj*sdd+ii*ps);
}
}
ii += 4;
}
// 1 2 3
if(ii<imax)
{
kernel_dgelqf_vs_lib4(m-ii, n-ii, imax-ii, ii&(ps-1), pD+ii*sdd+ii*ps, sdd, dD+ii);
}
#else // no haswell
for(ii=0; ii<imax-4; ii+=4)
{
// kernel_dgelqf_vs_lib4(4, n-ii, 4, 0, pD+ii*sdd+ii*ps, sdd, dD+ii);
// kernel_dgelqf_4_lib4(n-ii, pD+ii*sdd+ii*ps, dD+ii);
// kernel_dlarft_4_lib4(n-ii, pD+ii*sdd+ii*ps, dD+ii, pT);
kernel_dgelqf_dlarft4_4_lib4(n-ii, pD+ii*sdd+ii*ps, dD+ii, pT);
jj = ii+4;
#if defined(TARGET_X64_INTEL_SANDY_BRIDGE)
for(; jj<m-7; jj+=8)
{
kernel_dlarfb4_r_8_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps, sdd);
}
#endif
for(; jj<m-3; jj+=4)
{
kernel_dlarfb4_r_4_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+jj*sdd+ii*ps);
}
for(ll=0; ll<m-jj; ll++)
{
kernel_dlarfb4_r_1_lib4(n-ii, pD+ii*sdd+ii*ps, pT, pD+ll+jj*sdd+ii*ps);
}
}
if(ii<imax)
{
if(ii==imax-4)
{
kernel_dgelqf_4_lib4(n-ii, pD+ii*sdd+ii*ps, dD+ii);
}
else
{
kernel_dgelqf_vs_lib4(m-ii, n-ii, imax-ii, ii&(ps-1), pD+ii*sdd+ii*ps, sdd, dD+ii);
}
}
#endif // no haswell
#endif
return;
}
#else
#error : wrong LA choice
#endif