Sprinter/mame
2
0
Fork 0
mirror of https://github.com/holub/mame synced 2025-04-18 22:49:58 +03:00
Code Issues Actions Packages Projects Releases Wiki Activity

Increase performance from 53% to 69% for matrix size around 30x30. These

Browse Source 
matrices are e.g. used in congo bongo without optimisation. (nw)
This commit is contained in:
couriersud 2015-08-20 00:31:18 +02:00
parent eacd7ef4b2
commit 1d875f33fe
3 changed files with 39 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
nl_examples/congo_bongo.c
View File

@ -101,7 +101,7 @@ NETLIST_START(dummy)
PARAM(XU13.D.MODEL, "MB3614(TYPE=1)")
#endif
#if 0
#if 1
OPTIMIZE_FRONTIER(C51.1, RES_K(20), 50)
OPTIMIZE_FRONTIER(R77.2, RES_K(20), 50)

60
src/emu/netlist/solver/nld_ms_direct.h
View File

@ -265,6 +265,33 @@ ATTR_COLD void matrix_solver_direct_t<m_N, _storage_N>::vsetup(analog_net_t::lis
psort_list(t->m_nz);
}
/* create a list of non zero elements below diagonal k
* This should reduce cache misses ...
*/
bool touched[_storage_N][_storage_N] = { false };
for (unsigned k = 0; k < N(); k++)
{
m_terms[k]->m_nzbd.clear();
for (unsigned j = 0; j < m_terms[k]->m_nz.size(); j++)
touched[k][m_terms[k]->m_nz[j]] = true;
}
for (unsigned k = 0; k < N(); k++)
{
for (unsigned row = k + 1; row < N(); row++)
{
if (touched[row][k])
{
if (!m_terms[k]->m_nzbd.contains(row))
m_terms[k]->m_nzbd.add(row);
for (unsigned col = k; col < N(); col++)
if (touched[k][col])
touched[row][col] = true;
}
}
}
if (0)
for (unsigned k = 0; k < N(); k++)
{
@ -393,37 +420,22 @@ ATTR_HOT void matrix_solver_direct_t<m_N, _storage_N>::LE_solve()
{
/* FIXME: Singular matrix? */
const nl_double f = 1.0 / A(i,i);
const unsigned *p = m_terms[i]->m_nzrd.data();
const unsigned * RESTRICT const p = m_terms[i]->m_nzrd.data();
const unsigned e = m_terms[i]->m_nzrd.size();
/* Eliminate column i from row j */
for (unsigned j = i + 1; j < kN; j++)
const unsigned * RESTRICT const pb = m_terms[i]->m_nzbd.data();
const unsigned eb = m_terms[i]->m_nzbd.size();
for (unsigned jb = 0; jb < eb; jb++)
{
if (A(j,i) != NL_FCONST(0.0))
const unsigned j = pb[jb];
const nl_double f1 = - A(j,i) * f;
for (unsigned k = 0; k < e; k++)
{
const nl_double f1 = - A(j,i) * f;
#if 0
/* The code below is 30% faster than the original
* implementation which is given here for reference.
*
* for (unsigned k = i + 1; k < kN; k++)
* m_A[j][k] = m_A[j][k] + m_A[i][k] * f1;
*/
double * RESTRICT d = &m_A[j][i+1];
const double * RESTRICT s = &m_A[i][i+1];
const int e = kN - i - 1;
for (int k = 0; k < e; k++)
d[k] = d[k] + s[k] * f1;
#else
for (unsigned k = 0; k < e; k++)
{
const unsigned pk = p[k];
A(j,pk) += A(i,pk) * f1;
}
#endif
m_RHS[j] += m_RHS[i] * f1;
A(j,p[k]) += A(i,p[k]) * f1;
}
m_RHS[j] += m_RHS[i] * f1;
}
}
}

3
src/emu/netlist/solver/nld_solver.h
View File

@ -84,8 +84,9 @@ class terms_t
unsigned m_railstart;
plist_t<unsigned> m_nzrd; /* non zero right of the diagonal for elimination */
plist_t<unsigned> m_nz; /* all non zero for multiplication */
plist_t<unsigned> m_nzrd; /* non zero right of the diagonal for elimination */
plist_t<unsigned> m_nzbd; /* non zero below of the diagonal for elimination */
private:
plist_t<terminal_t *> m_term;
plist_t<int> m_net_other;