use omp simd instead of direct call to libmvec; add fallback for intrinsics if not supported
This commit is contained in:
Florian Ballenegger
parent
bbcd5baa84
commit
4c90028583
|
|
@ -55,136 +55,123 @@
|
|||
|
||||
#define USEX86INTRINSICS 1
|
||||
|
||||
typedef double Vec4d __attribute__ ((vector_size (sizeof(double)*NSIMD), aligned (sizeof(double)*NSIMD)));
|
||||
typedef long int Vec4m __attribute__ ((vector_size (sizeof(double)*NSIMD), aligned (sizeof(double)*NSIMD)));
|
||||
typedef double Vec4d __attribute__ ((vector_size (sizeof(double)*4), aligned (sizeof(double)*4)));
|
||||
typedef int64_t Vec4m __attribute__ ((vector_size (sizeof(double)*4), aligned (sizeof(double)*4)));
|
||||
|
||||
|
||||
#define SIMDANY(err) (err!=0)
|
||||
#define SIMDIFYCMD(cmd) /* empty */
|
||||
#define SIMDifySaveScope(sc) /* empty */
|
||||
|
||||
#define vec4_pow0p7(x,p) vec4_mypow(x,p)
|
||||
#define vec4_powMJ(x,p) vec4_mypow(x,p)
|
||||
#define vec4_powMJSW(x,p) vec4_mypow(x,p)
|
||||
#define vec4_powMJSWG(x,p) vec4_mypow(x,p)
|
||||
|
||||
#if USEX86INTRINSICS==1
|
||||
/* libmvec prototypes */
|
||||
/* Caution: those libmvec functions are not as precise as std libm */
|
||||
__m256d _ZGVdN4v_exp(__m256d x);
|
||||
__m256d _ZGVdN4v_log(__m256d x);
|
||||
|
||||
#define vec4_MAX(a,b) _mm256_max_pd(a,b)
|
||||
#define vec4_exp(a) _ZGVdN4v_exp(a)
|
||||
#define vec4_log(a) _ZGVdN4v_log(a)
|
||||
#define vec4_sqrt(a) _mm256_sqrt_pd(a)
|
||||
|
||||
|
||||
static inline Vec4d vec4_blend(Vec4d fa, Vec4d tr, Vec4m mask)
|
||||
{
|
||||
return _mm256_blendv_pd(fa,tr, (Vec4d) mask);
|
||||
}
|
||||
|
||||
static inline Vec4d vec4_fabs(Vec4d x)
|
||||
{
|
||||
return vec4_blend(x,-x,x<0);
|
||||
}
|
||||
|
||||
#else
|
||||
/* vector-libm prototypes */
|
||||
Vec4d vec4_exp_vectorlibm(Vec4d x); /* defined in vec4_exp.c */
|
||||
Vec4d vec4_log_vectorlibm(Vec4d x); /* defined in vec4_log.c */
|
||||
#define vec4_exp(a) vec4_exp_vectorlibm(a)
|
||||
#define vec4_log(a) vec4_log_vectorlibm(a)
|
||||
static inline Vec4d vec4_MAX(Vec4d a, Vec4d b)
|
||||
{
|
||||
return vec4_blend(a,b,a<b);
|
||||
}
|
||||
static inline Vec4d vec4_blend(Vec4d fa, Vec4d tr, Vec4m mask)
|
||||
{
|
||||
/* hope for good vectorization by the compiler ! */
|
||||
Vec4d res;
|
||||
Vec4d r;
|
||||
#pragma omp simd
|
||||
for(int i=0;i<4;i++)
|
||||
{
|
||||
res[i] = mask[i] ? tr[i] : fa[i];
|
||||
}
|
||||
return res;
|
||||
}
|
||||
static inline Vec4d vec4_fabs(Vec4d x)
|
||||
{
|
||||
/* hope for good vectorization by the compiler ! */
|
||||
Vec4d res;
|
||||
#pragma omp simd
|
||||
for(int i=0;i<4;i++)
|
||||
{
|
||||
res[i] = (x[i] < 0) ? -x[i] : x[i];
|
||||
}
|
||||
return res;
|
||||
}
|
||||
static inline Vec4d vec4_sqrt(Vec4d x)
|
||||
{
|
||||
/* hope for good vectorization by the compiler ! */
|
||||
Vec4d res;
|
||||
#pragma omp simd
|
||||
for(int i=0;i<4;i++)
|
||||
{
|
||||
res[i] = sqrt(x[i]);
|
||||
}
|
||||
return res;
|
||||
r[i] = (mask[i]==0 ? fa[i] : tr[i]);
|
||||
return r;
|
||||
}
|
||||
#endif
|
||||
|
||||
static inline Vec4d vec4_mypow(Vec4d x, double p)
|
||||
static inline Vec4d vec4_exp(Vec4d x)
|
||||
{
|
||||
Vec4d r;
|
||||
#pragma omp simd
|
||||
for(int i=0;i<4;i++)
|
||||
r[i] = exp(x[i]);
|
||||
return r;
|
||||
}
|
||||
|
||||
static inline Vec4d vec4_log(Vec4d x)
|
||||
{
|
||||
Vec4d r;
|
||||
#pragma omp simd
|
||||
for(int i=0;i<4;i++)
|
||||
r[i] = log(x[i]);
|
||||
return r;
|
||||
}
|
||||
|
||||
static inline Vec4d vec4_max(Vec4d x, Vec4d y)
|
||||
{
|
||||
Vec4d r;
|
||||
#pragma omp simd
|
||||
for(int i=0;i<4;i++)
|
||||
r[i] = MAX(x[i],y[i]);
|
||||
return r;
|
||||
}
|
||||
|
||||
static inline Vec4d vec4_sqrt(Vec4d x)
|
||||
{
|
||||
Vec4d r;
|
||||
#pragma omp simd
|
||||
for(int i=0;i<4;i++)
|
||||
r[i] = sqrt(x[i]);
|
||||
return r;
|
||||
}
|
||||
|
||||
static inline Vec4d vec4_fabs(Vec4d x)
|
||||
{
|
||||
Vec4d r;
|
||||
#pragma omp simd
|
||||
for(int i=0;i<4;i++)
|
||||
r[i] = fabs(x[i]);
|
||||
return r;
|
||||
}
|
||||
|
||||
#define vec4_pow0p7(x,p) vec4_pow(x,p)
|
||||
#define vec4_powMJ(x,p) vec4_pow(x,p)
|
||||
#define vec4_powMJSW(x,p) vec4_pow(x,p)
|
||||
#define vec4_powMJSWG(x,p) vec4_pow(x,p)
|
||||
|
||||
static inline Vec4d vec4_pow(Vec4d x, double p)
|
||||
{
|
||||
return vec4_exp(vec4_log(x)*p);
|
||||
}
|
||||
|
||||
|
||||
/* some debug utils functions */
|
||||
void vec4_printd(const char* msg, const char* name, Vec4d vecd)
|
||||
{
|
||||
printf("%s %s %g %g %g %g\n",msg,name,vecd[0],vecd[1],vecd[2],vecd[3]);
|
||||
}
|
||||
|
||||
void vec4_printm(const char* msg, const char* name, Vec4m vecm)
|
||||
{
|
||||
printf("%s %s %ld %ld %ld %ld\n",msg,name,vecm[0],vecm[1],vecm[2],vecm[3]);
|
||||
}
|
||||
|
||||
void vec4_CheckCollisions(Vec4m stateindexes, const char* msg)
|
||||
{
|
||||
for(int i=0;i<NSIMD;i++)
|
||||
for(int j=0;j<NSIMD;j++)
|
||||
if(i!=j)
|
||||
if(stateindexes[i]==stateindexes[j])
|
||||
{
|
||||
printf("%s, collisions %ld %ld %ld %ld!\n",msg,stateindexes[0],stateindexes[1],stateindexes[2],stateindexes[3]);
|
||||
raise(SIGINT);
|
||||
}
|
||||
}
|
||||
|
||||
/* useful vectorized functions */
|
||||
static inline Vec4d SIMDLOADDATA(int idx, double data[7][NSIMD])
|
||||
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};
|
||||
}
|
||||
/* TODO: prefix with vec4_ */
|
||||
static inline Vec4d SIMDLOADDATA(int idx, double data[7][4])
|
||||
{
|
||||
return (Vec4d) {data[idx][0],data[idx][1],data[idx][2],data[idx][3]};
|
||||
}
|
||||
|
||||
static inline Vec4d vec4_BSIM3v32_StateAccess(double* cktstate, Vec4m stateindexes)
|
||||
{
|
||||
return (Vec4d) {
|
||||
Vec4d r;
|
||||
#pragma omp simd
|
||||
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]]
|
||||
};
|
||||
};"*/
|
||||
}
|
||||
|
||||
|
||||
static inline void vec4_BSIM3v32_StateStore(double* cktstate, Vec4m stateindexes, Vec4d values)
|
||||
{
|
||||
if(0) vec4_CheckCollisions(stateindexes,"SateStore");
|
||||
for(int idx=0;idx<NSIMD;idx++)
|
||||
/*if(0) vec4_CheckCollisions(stateindexes,"SateStore");*/
|
||||
#pragma omp simd
|
||||
for(int idx=0;idx<4;idx++)
|
||||
{
|
||||
cktstate[stateindexes[idx]] = values[idx];
|
||||
}
|
||||
|
|
@ -192,8 +179,9 @@ static inline void vec4_BSIM3v32_StateStore(double* cktstate, Vec4m stateindexes
|
|||
|
||||
static inline void vec4_BSIM3v32_StateAdd(double* cktstate, Vec4m stateindexes, Vec4d values)
|
||||
{
|
||||
if(0) vec4_CheckCollisions(stateindexes,"StateAdd");
|
||||
for(int idx=0;idx<NSIMD;idx++)
|
||||
/*if(0) vec4_CheckCollisions(stateindexes,"StateAdd");*/
|
||||
#pragma omp simd
|
||||
for(int idx=0;idx<4;idx++)
|
||||
{
|
||||
cktstate[stateindexes[idx]] += values[idx];
|
||||
}
|
||||
|
|
@ -201,30 +189,14 @@ static inline void vec4_BSIM3v32_StateAdd(double* cktstate, Vec4m stateindexes,
|
|||
|
||||
static inline void vec4_BSIM3v32_StateSub(double* cktstate, Vec4m stateindexes, Vec4d values)
|
||||
{
|
||||
if(0) vec4_CheckCollisions(stateindexes,"StateSub");
|
||||
for(int idx=0;idx<NSIMD;idx++)
|
||||
/*if(0) vec4_CheckCollisions(stateindexes,"StateSub");*/
|
||||
#pragma omp simd
|
||||
for(int idx=0;idx<4;idx++)
|
||||
{
|
||||
cktstate[stateindexes[idx]] -= values[idx];
|
||||
}
|
||||
}
|
||||
|
||||
static inline Vec4d vec4_exp_seq(Vec4d val)
|
||||
{
|
||||
return (Vec4d) {exp(val[0]),exp(val[1]),exp(val[2]),exp(val[3])};
|
||||
}
|
||||
static inline Vec4d vec4_log_seq(Vec4d val)
|
||||
{
|
||||
return (Vec4d) {log(val[0]),log(val[1]),log(val[2]),log(val[3])};
|
||||
}
|
||||
static inline Vec4d vec4_sqrt_seq(Vec4d val)
|
||||
{
|
||||
return (Vec4d) {sqrt(val[0]),sqrt(val[1]),sqrt(val[2]),sqrt(val[3])};
|
||||
}
|
||||
static inline Vec4d vec4_MAX_seq(Vec4d a, Vec4d b)
|
||||
{
|
||||
return (Vec4d) {MAX(a[0],b[0]),MAX(a[1],b[1]),MAX(a[2],b[2]),MAX(a[3],b[3])};
|
||||
}
|
||||
|
||||
static inline int vec4_BSIM3v32_ACM_saturationCurrents
|
||||
(
|
||||
BSIM3v32model *model,
|
||||
|
|
@ -235,7 +207,7 @@ static inline int vec4_BSIM3v32_ACM_saturationCurrents
|
|||
{
|
||||
int error;
|
||||
double dsat,ssat;
|
||||
for(int idx=0;idx<NSIMD;idx++)
|
||||
for(int idx=0;idx<4;idx++)
|
||||
{
|
||||
error = BSIM3v32_ACM_saturationCurrents(
|
||||
model, heres[idx],
|
||||
|
|
@ -264,7 +236,7 @@ static inline int vec4_BSIM3v32_ACM_junctionCapacitances(
|
|||
int error;
|
||||
double areaDB,periDB,gateDB,areaSB,periSB,gateSB;
|
||||
|
||||
for(int idx=0;idx<NSIMD;idx++)
|
||||
for(int idx=0;idx<4;idx++)
|
||||
{
|
||||
error = BSIM3v32_ACM_junctionCapacitances(
|
||||
model, heres[idx],
|
||||
|
|
@ -290,8 +262,8 @@ static inline int vec4_BSIM3v32_ACM_junctionCapacitances(
|
|||
static inline int vec4_NIintegrate(CKTcircuit* ckt, double* geq, double *ceq, double zero, Vec4m chargestate)
|
||||
{
|
||||
int error;
|
||||
if (0) vec4_CheckCollisions(chargestate, "NIIntegrate");
|
||||
for(int idx=0;idx<NSIMD;idx++)
|
||||
/*if (0) vec4_CheckCollisions(chargestate, "NIIntegrate");*/
|
||||
for(int idx=0;idx<4;idx++)
|
||||
{
|
||||
error = NIintegrate(ckt,geq,ceq,zero,chargestate[idx]);
|
||||
if(error) return error;
|
||||
|
|
@ -303,15 +275,35 @@ static inline int vec4_SIMDCOUNT(Vec4m mask) {
|
|||
return (mask[0] ? 1 : 0) + (mask[1] ? 1 : 0) + (mask[2] ? 1 : 0) + (mask[3] ? 1 : 0);
|
||||
}
|
||||
|
||||
static inline Vec4d vec4_SIMDTOVECTOR(double val)
|
||||
|
||||
|
||||
/* some debug utils functions */
|
||||
void vec4_printd(const char* msg, const char* name, Vec4d vecd)
|
||||
{
|
||||
return (Vec4d) {val,val,val,val};
|
||||
}
|
||||
static inline Vec4m vec4_SIMDTOVECTORMASK(int val)
|
||||
{
|
||||
return (Vec4m) {val,val,val,val};
|
||||
printf("%s %s %g %g %g %g\n",msg,name,vecd[0],vecd[1],vecd[2],vecd[3]);
|
||||
}
|
||||
|
||||
void vec4_printm(const char* msg, const char* name, Vec4m vecm)
|
||||
{
|
||||
printf("%s %s %ld %ld %ld %ld\n",msg,name,vecm[0],vecm[1],vecm[2],vecm[3]);
|
||||
}
|
||||
|
||||
void vec4_CheckCollisions(Vec4m stateindexes, const char* msg)
|
||||
{
|
||||
for(int i=0;i<4;i++)
|
||||
for(int j=0;j<4;j++)
|
||||
if(i!=j)
|
||||
if(stateindexes[i]==stateindexes[j])
|
||||
{
|
||||
printf("%s, collisions %ld %ld %ld %ld!\n",msg,stateindexes[0],stateindexes[1],stateindexes[2],stateindexes[3]);
|
||||
raise(SIGINT);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
#include "b3v32ldsimd8d.c"
|
||||
|
||||
typedef float vecelem_t;
|
||||
|
||||
int BSIM3v32LoadSIMD(BSIM3v32instance **heres, CKTcircuit *ckt
|
||||
#ifndef USE_OMP
|
||||
|
|
@ -335,7 +327,7 @@ int BSIM3v32LoadSIMD(BSIM3v32instance **heres, CKTcircuit *ckt
|
|||
#pragma message "Use OMP SIMD8 version"
|
||||
#include "b3v32ldseq_simd8_omp.c"
|
||||
#else
|
||||
#include "b3v32ldseq_simd8.c"
|
||||
#include "b3v32ldseq_simd8d.c"
|
||||
#endif
|
||||
#else
|
||||
#error Unsupported value for NSIMD
|
||||
|
|
|
|||
Loading…
Reference in New Issue