luke
/
ngspice
mirror of https://git.code.sf.net/p/ngspice/ngspice
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

testing for libmvec in configure.ac and use it if available

This commit is contained in:
Florian Ballenegger 2020-07-29 19:13:43 +02:00
parent a4fd01b7af
commit 34ebb23b9d
2 changed files with 93 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
configure.ac
View File

@ -1228,7 +1228,29 @@ if test "x$enable_modsimd" = xyes; then
])
fi
# now adjust compiler flags
# test if we can use libmvec or equivalent
if test "x$enable_modsimd" = xyes; then
AC_MSG_CHECKING([vector math library])
AC_TRY_COMPILE([
#include <math.h>
typedef double Vec4d __attribute__ ((vector_size (4*sizeof(double)), aligned (4*sizeof(double))));
Vec4d _ZGVdN4v_exp(Vec4d x);
Vec4d _ZGVdN4v_log(Vec4d x);
],[
Vec4d v1,v2;
v1 = (Vec4d) {1.0,2.0,3.141592,4.0};
v2 = _ZGVdN4v_log(v1);
v1 = _ZGVdN4v_exp(v2);
],[
AC_MSG_RESULT([yes])
have_libmvec=yes
], [
AC_MSG_RESULT([no])
have_libmvec=no
])
fi
# adjust compiler flags for MODSIMD
if test "x$enable_modsimd" = xyes; then
SIMD_CPPFLAGS=""
SIMD_LDFLAGS=""
@ -1262,6 +1284,11 @@ if test "x$enable_modsimd" = xyes; then
AC_MSG_RESULT([Use x86 simd intrinsics])
fi
if test "x$have_libmvec" = xyes; then
AC_DEFINE([HAS_LIBMVEC], [1], [can call libmvec exp and log functions])
AC_MSG_RESULT([Use libmvec])
fi
AC_MSG_RESULT([MODSIMD feature enabled])
fi

86
src/spicelib/devices/bsim3v32simd/b3v32ldsimd4d.c
View File

@ -27,8 +27,17 @@
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
******************************************************************************/
#undef USE_OMPSIMD
/* disable omp simd for GCC, as it slow down a bit */
#if !defined(__GNUC__) || defined(__clang__) || defined(__INTEL_COMPILER)
#define USE_OMPSIMD
#endif
/* HAS_LIBMVEC defined from configure.ac */
/* USE_SERIAL_FORM can be defined but has no performance influence */
/******* vec4_blend *******/
#if USEX86INTRINSICS==1
static inline Vec4d vec4_blend(Vec4d fa, Vec4d tr, Vec4m mask)
{
@ -47,6 +56,17 @@ static inline Vec4d vec4_blend(Vec4d fa, Vec4d tr, Vec4m mask)
}
#endif
/******* vec4_exp, vec4_log *******/
#ifdef HAS_LIBMVEC
Vec4d _ZGVdN4v_exp(Vec4d x);
Vec4d _ZGVdN4v_log(Vec4d x);
/*Vec4d _ZGVdN4vv_pow(Vec4d x, Vec4d y);*/
#define vec4_exp(a) _ZGVdN4v_exp(a)
#define vec4_log(a) _ZGVdN4v_log(a)
#endif
#ifndef HAS_LIBMVEC
static inline Vec4d vec4_exp(Vec4d x)
{
Vec4d r;
@ -68,8 +88,14 @@ static inline Vec4d vec4_log(Vec4d x)
r[i] = log(x[i]);
return r;
}
#endif
static inline Vec4d vec4_max(Vec4d x, Vec4d y)
/******* vec4_MAX, vec4_sqrt *******/
#ifdef USEX86INTRINSICS
#define vec4_MAX(a,b) _mm256_max_pd(a,b)
#define vec4_sqrt(a) _mm256_sqrt_pd(a)
#else
static inline Vec4d vec4_MAX(Vec4d x, Vec4d y)
{
Vec4d r;
#ifdef USE_OMPSIMD
@ -90,7 +116,15 @@ static inline Vec4d vec4_sqrt(Vec4d x)
r[i] = sqrt(x[i]);
return r;
}
#endif
/******* vec4_fabs *******/
#ifdef USE_SERIAL_FORM
static inline Vec4d vec4_fabs(Vec4d x)
{
return vec4_blend(x,-x,x<0);
}
#else
static inline Vec4d vec4_fabs(Vec4d x)
{
Vec4d r;
@ -101,6 +135,7 @@ static inline Vec4d vec4_fabs(Vec4d x)
r[i] = fabs(x[i]);
return r;
}
#endif
#define vec4_pow0p7(x,p) vec4_pow(x,p)
#define vec4_powMJ(x,p) vec4_pow(x,p)
@ -109,10 +144,23 @@ static inline Vec4d vec4_fabs(Vec4d x)
static inline Vec4d vec4_pow(Vec4d x, double p)
{
/*return _ZGVdN4vv_pow(x,(Vec4d) {p,p,p,p});*/
return vec4_exp(vec4_log(x)*p);
}
/* useful vectorized functions */
/******* vec4_SIMDTOVECTOR, vec4_SIMDTOVECTORMASK *******/
#ifdef USE_SERIAL_FORM
static inline Vec4d vec4_SIMDTOVECTOR(double val)
{
return (Vec4d) {val,val,val,val};
}
static inline Vec4m vec4_SIMDTOVECTORMASK(int val)
{
return (Vec4m) {val,val,val,val};
}
#else
static inline Vec4d vec4_SIMDTOVECTOR(double val)
{
Vec4d r;
@ -134,23 +182,26 @@ static inline Vec4m vec4_SIMDTOVECTORMASK(int val)
r[i] = val;
return r;
}
#if 0
static inline Vec4d vec4_SIMDTOVECTOR(double val)
{
return (Vec4d) {val,val,val,val};
}
static inline Vec4m vec4_SIMDTOVECTORMASK(int val)
{
return (Vec4m) {val,val,val,val};
}
#endif
static inline Vec4d vec4_SIMDLOADDATA(int idx, double data[7][4])
{
return (Vec4d) {data[idx][0],data[idx][1],data[idx][2],data[idx][3]};
}
/******* vec4_BSIM3v32_StateAccess *******/
#ifdef USE_SERIAL_FORM
static inline Vec4d vec4_BSIM3v32_StateAccess(double* cktstate, Vec4m stateindexes)
{
return (Vec4d) {
cktstate[stateindexes[0]],
cktstate[stateindexes[1]],
cktstate[stateindexes[2]],
cktstate[stateindexes[3]]
};
}
#else
static inline Vec4d vec4_BSIM3v32_StateAccess(double* cktstate, Vec4m stateindexes)
{
Vec4d r;
@ -160,15 +211,8 @@ static inline Vec4d vec4_BSIM3v32_StateAccess(double* cktstate, Vec4m stateindex
for(int i=0;i<4;i++)
r[i] = cktstate[stateindexes[i]];
return r;
/*return (Vec4d) {
cktstate[stateindexes[0]],
cktstate[stateindexes[1]],
cktstate[stateindexes[2]],
cktstate[stateindexes[3]]
};"*/
}
#endif
static inline void vec4_BSIM3v32_StateStore(double* cktstate, Vec4m stateindexes, Vec4d values)
{