aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

workong on shuffle/rotate for double, there seems to be a bug in cvt2zmm cvt2hilo

Browse Source
This commit is contained in:
evghenii
2013-10-05 15:23:55 +03:00
parent 8b0fc558cb
commit 10223cfac3
1 changed files with 83 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
examples/intrinsics/knc-i1x16.h
View File

@@ -143,8 +143,14 @@ PRE_ALIGN(64) struct __vec16_f
struct PRE_ALIGN(128) __vec16_d
{
__m512d v1;
__m512d v2;
union {
__m512d v1;
__m512d v_hi;
};
union {
__m512d v2;
__m512d v_lo;
};
FORCEINLINE __vec16_d() : v1(_mm512_undefined_pd()), v2(_mm512_undefined_pd()) {}
FORCEINLINE __vec16_d(const __m512d _v1, const __m512d _v2) : v1(_v1), v2(_v2) {}
FORCEINLINE __vec16_d(const __vec16_d &o) : v1(o.v1), v2(o.v2) {}
@@ -158,6 +164,40 @@ struct PRE_ALIGN(128) __vec16_d
}
FORCEINLINE const double& operator[](const int i) const { return ((double*)this)[i]; }
FORCEINLINE double& operator[](const int i) { return ((double*)this)[i]; }
FORCEINLINE __vec16_d cvt2hilo() const
{
__m512i _hi, _lo;
_hi = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xFF00,
_mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0),
_mm512_castpd_si512(v1));
_hi = _mm512_mask_permutevar_epi32(_hi, 0x00FF,
_mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1),
_mm512_castpd_si512(v2));
_lo = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xFF00,
_mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1),
_mm512_castpd_si512(v1));
_lo = _mm512_mask_permutevar_epi32(_lo, 0x00FF,
_mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0),
_mm512_castpd_si512(v2));
return __vec16_d(_mm512_castsi512_pd(_hi), _mm512_castsi512_pd(_lo));
}
FORCEINLINE __vec16_d cvt2zmm() const
{
__m512i _v1, _v2;
_v1 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA,
_mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8),
_mm512_castpd_si512(v_hi));
_v1 = _mm512_mask_permutevar_epi32(_v1, 0x5555,
_mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8),
_mm512_castpd_si512(v_lo));
_v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA,
_mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0),
_mm512_castpd_si512(v_hi));
_v2 = _mm512_mask_permutevar_epi32(_v2, 0x5555,
_mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0),
_mm512_castpd_si512(v_lo));
return __vec16_d(_mm512_castsi512_pd(_v1), _mm512_castsi512_pd(_v2));
}
} POST_ALIGN(128);
struct PRE_ALIGN(128) __vec16_i64
@@ -1247,8 +1287,49 @@ static FORCEINLINE __vec16_d __broadcast_double(__vec16_d v, int index)
return ret;
}
#define CASTD2F(_v_, _v_hi_, _v_lo_) \
__vec16_f _v_hi_, _v_lo_; \
{ \
const __vec16_d v = _v_.cvt2hilo(); \
_v_hi_ = _mm512_castpd_ps(v.v_hi); \
_v_lo_ = _mm512_castpd_ps(v.v_lo); }
#define CASTF2D(_ret_hi_, _ret_lo_) \
__vec16_d(_mm512_castps_pd(_ret_hi_), _mm512_castps_pd(_ret_lo_)).cvt2zmm()
#if 0 /* knc::testme there appears to be no tests in ./tests for checking this functionality */
static FORCEINLINE __vec16_d __rotate_double(const __vec16_d _v, const int index)
{
// return _v; /* this one passes all tests , but most not */
CASTD2F(_v, v_hi, v_lo);
const __vec16_f ret_hi = __rotate_float(v_hi, index);
const __vec16_f ret_lo = __rotate_float(v_lo, index);
return CASTF2D(ret_hi, ret_lo);
}
#else
ROTATE(__vec16_d, double, double)
#endif
#if 0 /* knc::fails ./tests/shuffle2-4.ispc ./tests/shuffle2-5.ispc */
static FORCEINLINE __vec16_d __shuffle_double(__vec16_d _v, const __vec16_i32 index)
{
CASTD2F(_v, v_hi, v_lo);
const __vec16_f ret_hi = __shuffle_float(v_hi, index);
const __vec16_f ret_lo = __shuffle_float(v_lo, index);
return CASTF2D(ret_hi, ret_lo);
}
static FORCEINLINE __vec16_d __shuffle2_double(__vec16_d _v0, __vec16_d _v1, const __vec16_i32 index)
{
CASTD2F(_v0, v0_hi, v0_lo);
CASTD2F(_v1, v1_hi, v1_lo);
const __vec16_f ret_hi = __shuffle2_float(v0_hi, v1_hi, index);
const __vec16_f ret_lo = __shuffle2_float(v0_lo, v1_lo, index);
return CASTF2D(ret_hi, ret_lo);
}
#else
SHUFFLES(__vec16_d, double, double)
#endif
#undef CASTD2F
#undef CASTF2D
template <int ALIGN> static FORCEINLINE __vec16_d __load(const __vec16_d *p) \
{