#include <smmintrin.h>
|
#include <smmintrin.h>
|
#include <math.h>
|
#include <math.h>
|
|
|
#define NUM 64
|
#define NUM 64
|
|
|
static void
|
static void
|
init_round (FP_T *src)
|
init_round (FP_T *src)
|
{
|
{
|
int i, sign = 1;
|
int i, sign = 1;
|
FP_T f = rand ();
|
FP_T f = rand ();
|
|
|
for (i = 0; i < NUM; i++)
|
for (i = 0; i < NUM; i++)
|
{
|
{
|
src[i] = (i + 1)* f * M_PI * sign;
|
src[i] = (i + 1)* f * M_PI * sign;
|
if (i < (NUM / 2))
|
if (i < (NUM / 2))
|
{
|
{
|
if ((i % 6) == 0)
|
if ((i % 6) == 0)
|
f = f * src[i];
|
f = f * src[i];
|
}
|
}
|
else if (i == (NUM / 2))
|
else if (i == (NUM / 2))
|
f = rand ();
|
f = rand ();
|
else if ((i % 6) == 0)
|
else if ((i % 6) == 0)
|
f = 1 / (f * (i + 1) * src[i] * M_PI *sign);
|
f = 1 / (f * (i + 1) * src[i] * M_PI *sign);
|
sign = -sign;
|
sign = -sign;
|
}
|
}
|
}
|
}
|
|
|
static FP_T
|
static FP_T
|
do_round (FP_T f, int type)
|
do_round (FP_T f, int type)
|
{
|
{
|
short saved_cw, new_cw, clr_mask;
|
short saved_cw, new_cw, clr_mask;
|
FP_T ret;
|
FP_T ret;
|
|
|
if ((type & 4))
|
if ((type & 4))
|
{
|
{
|
type = 0;
|
type = 0;
|
clr_mask = 0xFFFF;
|
clr_mask = 0xFFFF;
|
}
|
}
|
else
|
else
|
{
|
{
|
type = 0x003F | ((type & 3) << 10);
|
type = 0x003F | ((type & 3) << 10);
|
clr_mask = ~0x0C3F;
|
clr_mask = ~0x0C3F;
|
}
|
}
|
|
|
__asm__ ("fld" ASM_SUFFIX " %0" : : "m" (*&f));
|
__asm__ ("fld" ASM_SUFFIX " %0" : : "m" (*&f));
|
|
|
__asm__ ("fstcw %0" : "=m" (*&saved_cw));
|
__asm__ ("fstcw %0" : "=m" (*&saved_cw));
|
new_cw = saved_cw & clr_mask;
|
new_cw = saved_cw & clr_mask;
|
new_cw |= type;
|
new_cw |= type;
|
__asm__ ("fldcw %0" : : "m" (*&new_cw));
|
__asm__ ("fldcw %0" : : "m" (*&new_cw));
|
|
|
__asm__ ("frndint\n"
|
__asm__ ("frndint\n"
|
"fstp" ASM_SUFFIX " %0\n" : "=m" (*&ret));
|
"fstp" ASM_SUFFIX " %0\n" : "=m" (*&ret));
|
__asm__ ("fldcw %0" : : "m" (*&saved_cw));
|
__asm__ ("fldcw %0" : : "m" (*&saved_cw));
|
return ret;
|
return ret;
|
}
|
}
|
|
|
static void
|
static void
|
sse4_1_test (void)
|
sse4_1_test (void)
|
{
|
{
|
int i;
|
int i;
|
FP_T f;
|
FP_T f;
|
union
|
union
|
{
|
{
|
VEC_T x[NUM / LOOP_INCREMENT];
|
VEC_T x[NUM / LOOP_INCREMENT];
|
FP_T f[NUM];
|
FP_T f[NUM];
|
} dst, src;
|
} dst, src;
|
|
|
init_round (src.f);
|
init_round (src.f);
|
|
|
for (i = 0; i < NUM / LOOP_INCREMENT; i++)
|
for (i = 0; i < NUM / LOOP_INCREMENT; i++)
|
dst.x[i] = ROUND_INTRIN (src.x[i], ROUND_MODE);
|
dst.x[i] = ROUND_INTRIN (src.x[i], ROUND_MODE);
|
|
|
for (i = 0; i < NUM; i += CHECK_LOOP_INCREMENT)
|
for (i = 0; i < NUM; i += CHECK_LOOP_INCREMENT)
|
{
|
{
|
f = do_round (src.f[i], CHECK_ROUND_MODE);
|
f = do_round (src.f[i], CHECK_ROUND_MODE);
|
if (f != dst.f[i])
|
if (f != dst.f[i])
|
abort ();
|
abort ();
|
}
|
}
|
|
|
if (_MM_FROUND_TO_NEAREST_INT != 0x00
|
if (_MM_FROUND_TO_NEAREST_INT != 0x00
|
|| _MM_FROUND_TO_NEG_INF != 0x01
|
|| _MM_FROUND_TO_NEG_INF != 0x01
|
|| _MM_FROUND_TO_POS_INF != 0x02
|
|| _MM_FROUND_TO_POS_INF != 0x02
|
|| _MM_FROUND_TO_ZERO != 0x03
|
|| _MM_FROUND_TO_ZERO != 0x03
|
|| _MM_FROUND_CUR_DIRECTION != 0x04
|
|| _MM_FROUND_CUR_DIRECTION != 0x04
|
|| _MM_FROUND_RAISE_EXC != 0x00
|
|| _MM_FROUND_RAISE_EXC != 0x00
|
|| _MM_FROUND_NO_EXC != 0x08
|
|| _MM_FROUND_NO_EXC != 0x08
|
|| _MM_FROUND_NINT != 0x00
|
|| _MM_FROUND_NINT != 0x00
|
|| _MM_FROUND_FLOOR != 0x01
|
|| _MM_FROUND_FLOOR != 0x01
|
|| _MM_FROUND_CEIL != 0x02
|
|| _MM_FROUND_CEIL != 0x02
|
|| _MM_FROUND_TRUNC != 0x03
|
|| _MM_FROUND_TRUNC != 0x03
|
|| _MM_FROUND_RINT != 0x04
|
|| _MM_FROUND_RINT != 0x04
|
|| _MM_FROUND_NEARBYINT != 0x0C)
|
|| _MM_FROUND_NEARBYINT != 0x0C)
|
abort ();
|
abort ();
|
}
|
}
|
|
|