| 1 |
/** |
| 2 |
* This file has no copyright assigned and is placed in the Public Domain. |
| 3 |
* This file is part of the mingw-w64 runtime package. |
| 4 |
* No warranty is given; refer to the file DISCLAIMER.PD within this package. |
| 5 |
*/ |
| 6 |
#ifndef _FVEC_H_INCLUDED |
| 7 |
#define _FVEC_H_INCLUDED |
| 8 |
|
| 9 |
#ifndef RC_INVOKED |
| 10 |
#ifndef __cplusplus |
| 11 |
#error ERROR: This file is only supported in C++ compilations! |
| 12 |
#endif |
| 13 |
|
| 14 |
#include <intrin.h> |
| 15 |
#include <assert.h> |
| 16 |
#include <crtdefs.h> |
| 17 |
|
| 18 |
#if defined(_ENABLE_VEC_DEBUG) |
| 19 |
#include <iostream> |
| 20 |
#endif |
| 21 |
|
| 22 |
#pragma pack(push,_CRT_PACKING) |
| 23 |
|
| 24 |
#ifdef __SSE__ |
| 25 |
|
| 26 |
#pragma pack(push,16) |
| 27 |
|
| 28 |
#define EXPLICIT explicit |
| 29 |
|
| 30 |
class F32vec4 { |
| 31 |
protected: |
| 32 |
__m128 vec; |
| 33 |
public: |
| 34 |
F32vec4() {} |
| 35 |
F32vec4(__m128 m) { vec = m;} |
| 36 |
F32vec4(float f3,float f2,float f1,float f0) { vec= _mm_set_ps(f3,f2,f1,f0); } |
| 37 |
EXPLICIT F32vec4(float f) { vec = _mm_set_ps1(f); } |
| 38 |
EXPLICIT F32vec4(double d) { vec = _mm_set_ps1((float) d); } |
| 39 |
F32vec4& operator =(float f) { vec = _mm_set_ps1(f); return *this; } |
| 40 |
F32vec4& operator =(double d) { vec = _mm_set_ps1((float) d); return *this; } |
| 41 |
operator __m128() const { return vec; } |
| 42 |
friend F32vec4 operator &(const F32vec4 &a,const F32vec4 &b) { return _mm_and_ps(a,b); } |
| 43 |
friend F32vec4 operator |(const F32vec4 &a,const F32vec4 &b) { return _mm_or_ps(a,b); } |
| 44 |
friend F32vec4 operator ^(const F32vec4 &a,const F32vec4 &b) { return _mm_xor_ps(a,b); } |
| 45 |
friend F32vec4 operator +(const F32vec4 &a,const F32vec4 &b) { return _mm_add_ps(a,b); } |
| 46 |
friend F32vec4 operator -(const F32vec4 &a,const F32vec4 &b) { return _mm_sub_ps(a,b); } |
| 47 |
friend F32vec4 operator *(const F32vec4 &a,const F32vec4 &b) { return _mm_mul_ps(a,b); } |
| 48 |
friend F32vec4 operator /(const F32vec4 &a,const F32vec4 &b) { return _mm_div_ps(a,b); } |
| 49 |
F32vec4& operator =(const F32vec4 &a) { vec = a.vec; return *this; } |
| 50 |
F32vec4& operator =(const __m128 &avec) { vec = avec; return *this; } |
| 51 |
F32vec4& operator +=(F32vec4 &a) { return *this = _mm_add_ps(vec,a); } |
| 52 |
F32vec4& operator -=(F32vec4 &a) { return *this = _mm_sub_ps(vec,a); } |
| 53 |
F32vec4& operator *=(F32vec4 &a) { return *this = _mm_mul_ps(vec,a); } |
| 54 |
F32vec4& operator /=(F32vec4 &a) { return *this = _mm_div_ps(vec,a); } |
| 55 |
F32vec4& operator &=(F32vec4 &a) { return *this = _mm_and_ps(vec,a); } |
| 56 |
F32vec4& operator |=(F32vec4 &a) { return *this = _mm_or_ps(vec,a); } |
| 57 |
F32vec4& operator ^=(F32vec4 &a) { return *this = _mm_xor_ps(vec,a); } |
| 58 |
friend float add_horizontal(F32vec4 &a) { |
| 59 |
F32vec4 ftemp = _mm_add_ss(a,_mm_add_ss(_mm_shuffle_ps(a,a,1),_mm_add_ss(_mm_shuffle_ps(a,a,2),_mm_shuffle_ps(a,a,3)))); |
| 60 |
return ftemp[0]; |
| 61 |
} |
| 62 |
friend F32vec4 sqrt(const F32vec4 &a) { return _mm_sqrt_ps(a); } |
| 63 |
friend F32vec4 rcp(const F32vec4 &a) { return _mm_rcp_ps(a); } |
| 64 |
friend F32vec4 rsqrt(const F32vec4 &a) { return _mm_rsqrt_ps(a); } |
| 65 |
friend F32vec4 rcp_nr(const F32vec4 &a) { |
| 66 |
F32vec4 Ra0 = _mm_rcp_ps(a); |
| 67 |
return _mm_sub_ps(_mm_add_ps(Ra0,Ra0),_mm_mul_ps(_mm_mul_ps(Ra0,a),Ra0)); |
| 68 |
} |
| 69 |
friend F32vec4 rsqrt_nr(const F32vec4 &a) { |
| 70 |
static const F32vec4 fvecf0pt5(0.5f); |
| 71 |
static const F32vec4 fvecf3pt0(3.0f); |
| 72 |
F32vec4 Ra0 = _mm_rsqrt_ps(a); |
| 73 |
return (fvecf0pt5 *Ra0) *(fvecf3pt0 - (a *Ra0) *Ra0); |
| 74 |
|
| 75 |
} |
| 76 |
#define Fvec32s4_COMP(op) friend F32vec4 cmp##op (const F32vec4 &a,const F32vec4 &b) { return _mm_cmp##op##_ps(a,b); } |
| 77 |
Fvec32s4_COMP(eq) |
| 78 |
Fvec32s4_COMP(lt) |
| 79 |
Fvec32s4_COMP(le) |
| 80 |
Fvec32s4_COMP(gt) |
| 81 |
Fvec32s4_COMP(ge) |
| 82 |
Fvec32s4_COMP(neq) |
| 83 |
Fvec32s4_COMP(nlt) |
| 84 |
Fvec32s4_COMP(nle) |
| 85 |
Fvec32s4_COMP(ngt) |
| 86 |
Fvec32s4_COMP(nge) |
| 87 |
#undef Fvec32s4_COMP |
| 88 |
|
| 89 |
friend F32vec4 simd_min(const F32vec4 &a,const F32vec4 &b) { return _mm_min_ps(a,b); } |
| 90 |
friend F32vec4 simd_max(const F32vec4 &a,const F32vec4 &b) { return _mm_max_ps(a,b); } |
| 91 |
|
| 92 |
#if defined(_ENABLE_VEC_DEBUG) |
| 93 |
friend std::ostream & operator<<(std::ostream & os,const F32vec4 &a) { |
| 94 |
float *fp = (float*)&a; |
| 95 |
os << "[3]:" << *(fp+3) |
| 96 |
<< " [2]:" << *(fp+2) |
| 97 |
<< " [1]:" << *(fp+1) |
| 98 |
<< " [0]:" << *fp; |
| 99 |
return os; |
| 100 |
} |
| 101 |
#endif |
| 102 |
const float& operator[](int i) const { |
| 103 |
assert((0 <= i) && (i <= 3)); |
| 104 |
float *fp = (float*)&vec; |
| 105 |
return *(fp+i); |
| 106 |
} |
| 107 |
float& operator[](int i) { |
| 108 |
assert((0 <= i) && (i <= 3)); |
| 109 |
float *fp = (float*)&vec; |
| 110 |
return *(fp+i); |
| 111 |
} |
| 112 |
}; |
| 113 |
|
| 114 |
inline F32vec4 unpack_low(const F32vec4 &a,const F32vec4 &b) { return _mm_unpacklo_ps(a,b); } |
| 115 |
inline F32vec4 unpack_high(const F32vec4 &a,const F32vec4 &b) { return _mm_unpackhi_ps(a,b); } |
| 116 |
inline int move_mask(const F32vec4 &a) { return _mm_movemask_ps(a); } |
| 117 |
inline void loadu(F32vec4 &a,float *p) { a = _mm_loadu_ps(p); } |
| 118 |
inline void storeu(float *p,const F32vec4 &a) { _mm_storeu_ps(p,a); } |
| 119 |
inline void store_nta(float *p,F32vec4 &a) { _mm_stream_ps(p,a); } |
| 120 |
|
| 121 |
#define Fvec32s4_SELECT(op) inline F32vec4 select_##op (const F32vec4 &a,const F32vec4 &b,const F32vec4 &c,const F32vec4 &d) { F32vec4 mask = _mm_cmp##op##_ps(a,b); return((mask & c) | F32vec4((_mm_andnot_ps(mask,d)))); } |
| 122 |
Fvec32s4_SELECT(eq) |
| 123 |
Fvec32s4_SELECT(lt) |
| 124 |
Fvec32s4_SELECT(le) |
| 125 |
Fvec32s4_SELECT(gt) |
| 126 |
Fvec32s4_SELECT(ge) |
| 127 |
Fvec32s4_SELECT(neq) |
| 128 |
Fvec32s4_SELECT(nlt) |
| 129 |
Fvec32s4_SELECT(nle) |
| 130 |
Fvec32s4_SELECT(ngt) |
| 131 |
Fvec32s4_SELECT(nge) |
| 132 |
#undef Fvec32s4_SELECT |
| 133 |
|
| 134 |
#if 0 /* Commented until required types are defined */ |
| 135 |
inline Is16vec4 simd_max(const Is16vec4 &a,const Is16vec4 &b) { return _m_pmaxsw(a,b); } |
| 136 |
inline Is16vec4 simd_min(const Is16vec4 &a,const Is16vec4 &b) { return _m_pminsw(a,b); } |
| 137 |
inline Iu8vec8 simd_max(const Iu8vec8 &a,const Iu8vec8 &b) { return _m_pmaxub(a,b); } |
| 138 |
inline Iu8vec8 simd_min(const Iu8vec8 &a,const Iu8vec8 &b) { return _m_pminub(a,b); } |
| 139 |
inline Iu16vec4 simd_avg(const Iu16vec4 &a,const Iu16vec4 &b) { return _m_pavgw(a,b); } |
| 140 |
inline Iu8vec8 simd_avg(const Iu8vec8 &a,const Iu8vec8 &b) { return _m_pavgb(a,b); } |
| 141 |
inline int move_mask(const I8vec8 &a) { return _m_pmovmskb(a); } |
| 142 |
inline Iu16vec4 mul_high(const Iu16vec4 &a,const Iu16vec4 &b) { return _m_pmulhuw(a,b); } |
| 143 |
inline void mask_move(const I8vec8 &a,const I8vec8 &b,char *addr) { _m_maskmovq(a,b,addr); } |
| 144 |
inline void store_nta(__m64 *p,M64 &a) { _mm_stream_pi(p,a); } |
| 145 |
inline int F32vec4ToInt(const F32vec4 &a) { return _mm_cvtt_ss2si(a); } |
| 146 |
inline Is32vec2 F32vec4ToIs32vec2 (const F32vec4 &a) { |
| 147 |
__m64 result; |
| 148 |
result = _mm_cvtt_ps2pi(a); |
| 149 |
return Is32vec2(result); |
| 150 |
} |
| 151 |
#endif |
| 152 |
|
| 153 |
inline F32vec4 IntToF32vec4(const F32vec4 &a,int i) { |
| 154 |
__m128 result; |
| 155 |
result = _mm_cvt_si2ss(a,i); |
| 156 |
return F32vec4(result); |
| 157 |
} |
| 158 |
|
| 159 |
#if 0 /* Commented until required types are defined */ |
| 160 |
inline F32vec4 Is32vec2ToF32vec4(const F32vec4 &a,const Is32vec2 &b) { |
| 161 |
__m128 result; |
| 162 |
result = _mm_cvt_pi2ps(a,b); |
| 163 |
return F32vec4(result); |
| 164 |
} |
| 165 |
#endif |
| 166 |
|
| 167 |
class F32vec1 { |
| 168 |
protected: |
| 169 |
__m128 vec; |
| 170 |
public: |
| 171 |
F32vec1() {} |
| 172 |
F32vec1(int i) { vec = _mm_cvt_si2ss(vec,i);}; |
| 173 |
EXPLICIT F32vec1(float f) { vec = _mm_set_ss(f); } |
| 174 |
EXPLICIT F32vec1(double d) { vec = _mm_set_ss((float) d); } |
| 175 |
F32vec1(__m128 m) { vec = m; } |
| 176 |
operator __m128() const { return vec; } |
| 177 |
friend F32vec1 operator &(const F32vec1 &a,const F32vec1 &b) { return _mm_and_ps(a,b); } |
| 178 |
friend F32vec1 operator |(const F32vec1 &a,const F32vec1 &b) { return _mm_or_ps(a,b); } |
| 179 |
friend F32vec1 operator ^(const F32vec1 &a,const F32vec1 &b) { return _mm_xor_ps(a,b); } |
| 180 |
friend F32vec1 operator +(const F32vec1 &a,const F32vec1 &b) { return _mm_add_ss(a,b); } |
| 181 |
friend F32vec1 operator -(const F32vec1 &a,const F32vec1 &b) { return _mm_sub_ss(a,b); } |
| 182 |
friend F32vec1 operator *(const F32vec1 &a,const F32vec1 &b) { return _mm_mul_ss(a,b); } |
| 183 |
friend F32vec1 operator /(const F32vec1 &a,const F32vec1 &b) { return _mm_div_ss(a,b); } |
| 184 |
F32vec1& operator +=(F32vec1 &a) { return *this = _mm_add_ss(vec,a); } |
| 185 |
F32vec1& operator -=(F32vec1 &a) { return *this = _mm_sub_ss(vec,a); } |
| 186 |
F32vec1& operator *=(F32vec1 &a) { return *this = _mm_mul_ss(vec,a); } |
| 187 |
F32vec1& operator /=(F32vec1 &a) { return *this = _mm_div_ss(vec,a); } |
| 188 |
F32vec1& operator &=(F32vec1 &a) { return *this = _mm_and_ps(vec,a); } |
| 189 |
F32vec1& operator |=(F32vec1 &a) { return *this = _mm_or_ps(vec,a); } |
| 190 |
F32vec1& operator ^=(F32vec1 &a) { return *this = _mm_xor_ps(vec,a); } |
| 191 |
friend F32vec1 sqrt(const F32vec1 &a) { return _mm_sqrt_ss(a); } |
| 192 |
friend F32vec1 rcp(const F32vec1 &a) { return _mm_rcp_ss(a); } |
| 193 |
friend F32vec1 rsqrt(const F32vec1 &a) { return _mm_rsqrt_ss(a); } |
| 194 |
friend F32vec1 rcp_nr(const F32vec1 &a) { |
| 195 |
F32vec1 Ra0 = _mm_rcp_ss(a); |
| 196 |
return _mm_sub_ss(_mm_add_ss(Ra0,Ra0),_mm_mul_ss(_mm_mul_ss(Ra0,a),Ra0)); |
| 197 |
} |
| 198 |
friend F32vec1 rsqrt_nr(const F32vec1 &a) { |
| 199 |
static const F32vec1 fvecf0pt5(0.5f); |
| 200 |
static const F32vec1 fvecf3pt0(3.0f); |
| 201 |
F32vec1 Ra0 = _mm_rsqrt_ss(a); |
| 202 |
return (fvecf0pt5 *Ra0) *(fvecf3pt0 - (a *Ra0) *Ra0); |
| 203 |
} |
| 204 |
#define Fvec32s1_COMP(op) friend F32vec1 cmp##op (const F32vec1 &a,const F32vec1 &b) { return _mm_cmp##op##_ss(a,b); } |
| 205 |
Fvec32s1_COMP(eq) |
| 206 |
Fvec32s1_COMP(lt) |
| 207 |
Fvec32s1_COMP(le) |
| 208 |
Fvec32s1_COMP(gt) |
| 209 |
Fvec32s1_COMP(ge) |
| 210 |
Fvec32s1_COMP(neq) |
| 211 |
Fvec32s1_COMP(nlt) |
| 212 |
Fvec32s1_COMP(nle) |
| 213 |
Fvec32s1_COMP(ngt) |
| 214 |
Fvec32s1_COMP(nge) |
| 215 |
#undef Fvec32s1_COMP |
| 216 |
|
| 217 |
friend F32vec1 simd_min(const F32vec1 &a,const F32vec1 &b) { return _mm_min_ss(a,b); } |
| 218 |
friend F32vec1 simd_max(const F32vec1 &a,const F32vec1 &b) { return _mm_max_ss(a,b); } |
| 219 |
|
| 220 |
#if defined(_ENABLE_VEC_DEBUG) |
| 221 |
friend std::ostream & operator<<(std::ostream & os,const F32vec1 &a) { |
| 222 |
float *fp = (float*)&a; |
| 223 |
os << "float:" << *fp; |
| 224 |
return os; |
| 225 |
} |
| 226 |
#endif |
| 227 |
}; |
| 228 |
|
| 229 |
#define Fvec32s1_SELECT(op) inline F32vec1 select_##op (const F32vec1 &a,const F32vec1 &b,const F32vec1 &c,const F32vec1 &d) { F32vec1 mask = _mm_cmp##op##_ss(a,b); return((mask & c) | F32vec1((_mm_andnot_ps(mask,d)))); } |
| 230 |
Fvec32s1_SELECT(eq) |
| 231 |
Fvec32s1_SELECT(lt) |
| 232 |
Fvec32s1_SELECT(le) |
| 233 |
Fvec32s1_SELECT(gt) |
| 234 |
Fvec32s1_SELECT(ge) |
| 235 |
Fvec32s1_SELECT(neq) |
| 236 |
Fvec32s1_SELECT(nlt) |
| 237 |
Fvec32s1_SELECT(nle) |
| 238 |
Fvec32s1_SELECT(ngt) |
| 239 |
Fvec32s1_SELECT(nge) |
| 240 |
#undef Fvec32s1_SELECT |
| 241 |
|
| 242 |
inline int F32vec1ToInt(const F32vec1 &a) |
| 243 |
{ |
| 244 |
return _mm_cvtt_ss2si(a); |
| 245 |
} |
| 246 |
|
| 247 |
#pragma pack(pop) |
| 248 |
|
| 249 |
#endif /* #ifdef __SSE__ */ |
| 250 |
#pragma pack(pop) |
| 251 |
|
| 252 |
#include <ivec.h> |
| 253 |
|
| 254 |
#endif |
| 255 |
#endif |
