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