void
run_one (FILE *fp, struct speed_params *s, mp_size_t prev_size)
{
const char *first_open_fastest, *first_open_notfastest, *first_close;
int j,i, fastest, want_data;
double fastest_time;
TMP_DECL;
TMP_MARK;
/* allocate data, unless all routines are NODATA */
want_data = 0;
for (i = 0; i < num_choices; i++)
want_data |= ((choice[i].p->flag & FLAG_NODATA) == 0);
if (want_data)
{
SPEED_TMP_ALLOC_LIMBS (sp.xp, s->size, s->align_xp);
SPEED_TMP_ALLOC_LIMBS (sp.yp, s->size, s->align_yp);
data_fill (s->xp, s->size);
data_fill (s->yp, s->size);
}
else
{
sp.xp = NULL;
sp.yp = NULL;
}
if (prev_size == -1 && option_cmp == CMP_DIFFPREV)
{
first_open_fastest = "(#";
first_open_notfastest = " (";
first_close = ")";
}
else
{
first_open_fastest = "#";
first_open_notfastest = " ";
first_close = "";
}
fastest = -1;
fastest_time = -1.0;
for (i = 0; i < num_choices; i++)
{
if( choice[i].nsum!=0)continue;
s->r = choice[i].r;
if( choice[i].colfile==-1)
{choice[i].time = speed_measure (choice[i].p->fun, s);}
else
{FILE *fp;char buf[1024],buf2[1024],*p;int got=0;
choice[i].time=-1.0;
fp=fopen(choice[i].filename,"rt");
if(fp==0){printf("Cant open %s\n",choice[i].filename);exit(1);}
while(fgets(buf,1024,fp)!=0)
if(atoi(buf)==s->size)
{p=buf;
for(j=0;j<=choice[i].colfile;j++)
{if(sscanf(p," %s",buf2)!=1)break;
p=strstr(p,buf2)+strlen(buf2);}
if(j==choice[i].colfile+1)
{while((p=strstr(buf2,"#"))!=0)*p=' ';// exclude #
choice[i].time=atof(buf2);
}
break;}
fclose(fp);
}
choice[i].no_time = (choice[i].time == -1.0);
if (! choice[i].no_time)
choice[i].time *= choice[i].scale;
/* Apply the effect of CMP_DIFFPREV, but the new choice[i].prev_time
is before any differences. */
if(choice[i].colfile==-1)
{
double t;
t = choice[i].time;
if (t != -1.0 && option_cmp == CMP_DIFFPREV && prev_size != -1)
{
if (choice[i].prev_time == -1.0)
choice[i].no_time = 1;
else
choice[i].time = choice[i].time - choice[i].prev_time;
}
choice[i].prev_time = t;
}
}
for (i = 0; i < num_choices; i++)
{if(choice[i].nsum==0)continue;
choice[i].time=0;choice[i].no_time=0;
for(j=0;j<choice[i].nsum;j++)
{choice[i].time+=choice[choice[i].sum[j]].time;
if(choice[choice[i].sum[j]].no_time)choice[i].no_time=1;
}
}
for (i = 0; i < num_choices; i++)
{
if (choice[i].no_time || choice[i].colfile!=-1)
continue;
if (option_cmp == CMP_DIFFPREV)
{
/* Conversion for UNIT_CYCLESPERLIMB differs in CMP_DIFFPREV. */
if (option_unit == UNIT_CYCLES)
choice[i].time /= speed_cycletime;
else if (option_unit == UNIT_CYCLESPERLIMB)
{
if (prev_size == -1)
choice[i].time /= speed_cycletime;
else
choice[i].time /= (speed_cycletime
* (SIZE_TO_DIVISOR(s->size)
- SIZE_TO_DIVISOR(prev_size)));
}
}
else
{
if (option_unit == UNIT_CYCLES)
choice[i].time /= speed_cycletime;
else if (option_unit == UNIT_CYCLESPERLIMB)
choice[i].time /= (speed_cycletime * SIZE_TO_DIVISOR(s->size));
}
}
for (i = 0; i < num_choices; i++)
{
if (choice[i].no_time)
continue;
/* Look for the fastest after CMP_DIFFPREV has been applied, but
before CMP_RATIO or CMP_DIFFERENCE. There's only a fastest shown
if there's more than one routine. */
for(j=0;j<xcoln;j++)if(xcol[j]==i)break; // excluded from fastest choice
if (j==xcoln && num_choices > 1 && (fastest == -1 || choice[i].time < fastest_time))
{
fastest = i;
fastest_time = choice[i].time;
}
}
for (i = 0; i < num_choices; i++)
{
if (choice[i].no_time )
continue;
if (option_cmp != CMP_DIFFPREV)
{
if (option_cmp == CMP_RATIO && i != option_cmp_pos)
{
/* A ratio isn't affected by the units chosen. */
if (choice[option_cmp_pos].no_time || choice[option_cmp_pos].time == 0.0)
choice[i].no_time = 1;
else
choice[i].time /= choice[option_cmp_pos].time;
}
else if (option_cmp == CMP_DIFFERENCE && i != option_cmp_pos)
{
if (choice[option_cmp_pos].no_time)
{
choice[i].no_time = 1;
continue;
}
choice[i].time -= choice[option_cmp_pos].time;
}
}
}
if (option_gnuplot)
{
/* In CMP_DIFFPREV, don't print anything for the first size, start
with the second where an actual difference is available.
In CMP_RATIO, print the "first" ie option_cmp_pos column as 1.0.
The 9 decimals printed is much more than the expected precision of
the measurements actually. */
if (! (option_cmp == CMP_DIFFPREV && prev_size == -1))
{
fprintf (fp, "%-6ld ", s->size);
for (i = 0; i < num_choices; i++)
fprintf (fp, " %.9e",
choice[i].no_time ? 0.0
: (option_cmp == CMP_RATIO && i == option_cmp_pos) ? 1.0
: choice[i].time);
fprintf (fp, "\n");
}
}
else
{
fprintf (fp, "%-6ld ", s->size);
for (i = 0; i < num_choices; i++)
{
char buf[128];
int decimals;
if (choice[i].no_time)
{
fprintf (fp, " %*s", COLUMN_WIDTH, "n/a");
}
else
{if (option_unit == UNIT_CYCLESPERLIMB
|| (option_cmp == CMP_RATIO && i != option_cmp_pos))
decimals = 4;
else if (option_unit == UNIT_CYCLES)
decimals = 2;
else
decimals = 9;
sprintf (buf, "%s%.*f%s",
i == fastest ? first_open_fastest : first_open_notfastest,
decimals, choice[i].time, first_close);
fprintf (fp, " %*s", COLUMN_WIDTH, buf);
}
}
fprintf (fp, "\n");
}
TMP_FREE;
}
void
run_one (FILE *fp, struct speed_params *s, mp_size_t prev_size)
{
const char *first_open_fastest, *first_open_notfastest, *first_close;
int i, fastest, want_data;
double fastest_time;
TMP_DECL;
TMP_MARK;
/* allocate data, unless all routines are NODATA */
want_data = 0;
for (i = 0; i < num_choices; i++)
want_data |= ((choice[i].p->flag & FLAG_NODATA) == 0);
if (want_data)
{
SPEED_TMP_ALLOC_LIMBS (sp.xp, s->size, s->align_xp);
SPEED_TMP_ALLOC_LIMBS (sp.yp, s->size, s->align_yp);
data_fill (s->xp, s->size);
data_fill (s->yp, s->size);
}
else
{
sp.xp = NULL;
sp.yp = NULL;
}
if (prev_size == -1 && option_cmp == CMP_DIFFPREV)
{
first_open_fastest = "(#";
first_open_notfastest = " (";
first_close = ")";
}
else
{
first_open_fastest = "#";
first_open_notfastest = " ";
first_close = "";
}
fastest = -1;
fastest_time = -1.0;
for (i = 0; i < num_choices; i++)
{
s->r = choice[i].r;
choice[i].time = speed_measure (choice[i].p->fun, s);
choice[i].no_time = (choice[i].time == -1.0);
if (! choice[i].no_time)
choice[i].time *= choice[i].scale;
/* Apply the effect of CMP_DIFFPREV, but the new choice[i].prev_time
is before any differences. */
{
double t;
t = choice[i].time;
if (t != -1.0 && option_cmp == CMP_DIFFPREV && prev_size != -1)
{
if (choice[i].prev_time == -1.0)
choice[i].no_time = 1;
else
choice[i].time = choice[i].time - choice[i].prev_time;
}
choice[i].prev_time = t;
}
if (choice[i].no_time)
continue;
/* Look for the fastest after CMP_DIFFPREV has been applied, but
before CMP_RATIO or CMP_DIFFERENCE. There's only a fastest shown
if there's more than one routine. */
if (num_choices > 1 && (fastest == -1 || choice[i].time < fastest_time))
{
fastest = i;
fastest_time = choice[i].time;
}
if (option_cmp == CMP_DIFFPREV)
{
/* Conversion for UNIT_CYCLESPERLIMB differs in CMP_DIFFPREV. */
if (option_unit == UNIT_CYCLES)
choice[i].time /= speed_cycletime;
else if (option_unit == UNIT_CYCLESPERLIMB)
{
if (prev_size == -1)
choice[i].time /= speed_cycletime;
else
choice[i].time /= (speed_cycletime
* (SIZE_TO_DIVISOR(s->size)
- SIZE_TO_DIVISOR(prev_size)));
}
}
else
{
if (option_unit == UNIT_CYCLES)
choice[i].time /= speed_cycletime;
else if (option_unit == UNIT_CYCLESPERLIMB)
choice[i].time /= (speed_cycletime * SIZE_TO_DIVISOR(s->size));
if (option_cmp == CMP_RATIO && i > 0)
{
/* A ratio isn't affected by the units chosen. */
if (choice[0].no_time || choice[0].time == 0.0)
choice[i].no_time = 1;
else
choice[i].time /= choice[0].time;
}
else if (option_cmp == CMP_DIFFERENCE && i > 0)
{
if (choice[0].no_time)
{
choice[i].no_time = 1;
continue;
}
choice[i].time -= choice[0].time;
}
}
}
if (option_gnuplot)
{
/* In CMP_DIFFPREV, don't print anything for the first size, start
with the second where an actual difference is available.
In CMP_RATIO, print the first column as 1.0.
The 9 decimals printed is much more than the expected precision of
the measurements actually. */
if (! (option_cmp == CMP_DIFFPREV && prev_size == -1))
{
fprintf (fp, "%-6ld ", s->size);
for (i = 0; i < num_choices; i++)
fprintf (fp, " %.9e",
choice[i].no_time ? 0.0
: (option_cmp == CMP_RATIO && i == 0) ? 1.0
: choice[i].time);
fprintf (fp, "\n");
}
}
else
{
fprintf (fp, "%-6ld ", s->size);
for (i = 0; i < num_choices; i++)
{
char buf[128];
int decimals;
if (choice[i].no_time)
{
fprintf (fp, " %*s", COLUMN_WIDTH, "n/a");
}
else
{if (option_unit == UNIT_CYCLESPERLIMB
|| (option_cmp == CMP_RATIO && i > 0))
decimals = 4;
else if (option_unit == UNIT_CYCLES)
decimals = 2;
else
decimals = 9;
sprintf (buf, "%s%.*f%s",
i == fastest ? first_open_fastest : first_open_notfastest,
decimals, choice[i].time, first_close);
fprintf (fp, " %*s", COLUMN_WIDTH, buf);
}
}
fprintf (fp, "\n");
}
TMP_FREE;
}
