|
|
|
| 1 |
|
/** |
| 2 |
|
*** Copyright (C) 1985-1999 Intel Corporation. All rights reserved. |
| 3 |
|
*** |
| 4 |
|
*** The information and source code contained herein is the exclusive |
| 5 |
|
*** property of Intel Corporation and may not be disclosed, examined |
| 6 |
|
*** or reproduced in whole or in part without explicit written authorization |
| 7 |
|
*** from the company. |
| 8 |
|
*** |
| 9 |
|
**/ |
| 10 |
|
|
| 11 |
|
/* |
| 12 |
|
* Definition of a C++ class interface to Streaming SIMD Extension intrinsics. |
| 13 |
|
* |
| 14 |
|
* |
| 15 |
|
* File name : fvec.h Fvec class definitions |
| 16 |
|
* |
| 17 |
|
* Concept: A C++ abstraction of Streaming SIMD Extensions designed to improve |
| 18 |
|
* |
| 19 |
|
* programmer productivity. Speed and accuracy are sacrificed for utility. |
| 20 |
|
* |
| 21 |
|
* Facilitates an easy transition to compiler intrinsics |
| 22 |
|
* |
| 23 |
|
* or assembly language. |
| 24 |
|
* |
| 25 |
|
* F32vec4: 4 packed single precision |
| 26 |
|
* 32-bit floating point numbers |
| 27 |
|
*/ |
| 28 |
|
|
| 29 |
|
#ifndef _FVEC_H_INCLUDED |
| 30 |
|
#define _FVEC_H_INCLUDED |
| 31 |
|
#ifndef RC_INVOKED |
| 32 |
|
|
| 33 |
|
#if !defined __cplusplus |
| 34 |
|
#error ERROR: This file is only supported in C++ compilations! |
| 35 |
|
#endif /* !__cplusplus */ |
| 36 |
|
|
| 37 |
|
#if defined(_M_CEE_PURE) |
| 38 |
|
#error ERROR: This file is not supported in the pure mode! |
| 39 |
|
#else |
| 40 |
|
|
| 41 |
|
#include <xmmintrin.h> /* Streaming SIMD Extensions Intrinsics include file */ |
| 42 |
|
#include <assert.h> |
| 43 |
|
#include <ivec.h> |
| 44 |
|
#include <crtdefs.h> |
| 45 |
|
|
| 46 |
|
/* Define _ENABLE_VEC_DEBUG to enable std::ostream inserters for debug output */ |
| 47 |
|
#if defined(_ENABLE_VEC_DEBUG) |
| 48 |
|
#include <iostream> |
| 49 |
|
#endif |
| 50 |
|
|
| 51 |
|
#ifdef _MSC_VER |
| 52 |
|
#pragma pack(push,_CRT_PACKING) |
| 53 |
|
#endif /* _MSC_VER */ |
| 54 |
|
|
| 55 |
|
#pragma pack(push,16) /* Must ensure class & union 16-B aligned */ |
| 56 |
|
|
| 57 |
|
#define EXPLICIT explicit |
| 58 |
|
|
| 59 |
|
class F32vec4 |
| 60 |
|
{ |
| 61 |
|
protected: |
| 62 |
|
__m128 vec; |
| 63 |
|
public: |
| 64 |
|
|
| 65 |
|
/* Constructors: __m128, 4 floats, 1 float */ |
| 66 |
|
F32vec4() {} |
| 67 |
|
|
| 68 |
|
/* initialize 4 SP FP with __m128 data type */ |
| 69 |
|
F32vec4(__m128 m) { vec = m;} |
| 70 |
|
|
| 71 |
|
/* initialize 4 SP FPs with 4 floats */ |
| 72 |
|
F32vec4(float f3, float f2, float f1, float f0) { vec= _mm_set_ps(f3,f2,f1,f0); } |
| 73 |
|
|
| 74 |
|
/* Explicitly initialize each of 4 SP FPs with same float */ |
| 75 |
|
EXPLICIT F32vec4(float f) { vec = _mm_set_ps1(f); } |
| 76 |
|
|
| 77 |
|
/* Explicitly initialize each of 4 SP FPs with same double */ |
| 78 |
|
EXPLICIT F32vec4(double d) { vec = _mm_set_ps1((float) d); } |
| 79 |
|
|
| 80 |
|
/* Assignment operations */ |
| 81 |
|
|
| 82 |
|
F32vec4& operator =(float f) { vec = _mm_set_ps1(f); return *this; } |
| 83 |
|
|
| 84 |
|
F32vec4& operator =(double d) { vec = _mm_set_ps1((float) d); return *this; } |
| 85 |
|
|
| 86 |
|
/* Conversion functions */ |
| 87 |
|
operator __m128() const { return vec; } /* Convert to __m128 */ |
| 88 |
|
|
| 89 |
|
/* Logical Operators */ |
| 90 |
|
friend F32vec4 operator &(const F32vec4 &a, const F32vec4 &b) { return _mm_and_ps(a,b); } |
| 91 |
|
friend F32vec4 operator |(const F32vec4 &a, const F32vec4 &b) { return _mm_or_ps(a,b); } |
| 92 |
|
friend F32vec4 operator ^(const F32vec4 &a, const F32vec4 &b) { return _mm_xor_ps(a,b); } |
| 93 |
|
|
| 94 |
|
/* Arithmetic Operators */ |
| 95 |
|
friend F32vec4 operator +(const F32vec4 &a, const F32vec4 &b) { return _mm_add_ps(a,b); } |
| 96 |
|
friend F32vec4 operator -(const F32vec4 &a, const F32vec4 &b) { return _mm_sub_ps(a,b); } |
| 97 |
|
friend F32vec4 operator *(const F32vec4 &a, const F32vec4 &b) { return _mm_mul_ps(a,b); } |
| 98 |
|
friend F32vec4 operator /(const F32vec4 &a, const F32vec4 &b) { return _mm_div_ps(a,b); } |
| 99 |
|
|
| 100 |
|
F32vec4& operator =(const F32vec4 &a) { vec = a.vec; return *this; } |
| 101 |
|
F32vec4& operator =(const __m128 &avec) { vec = avec; return *this; } |
| 102 |
|
F32vec4& operator +=(F32vec4 &a) { return *this = _mm_add_ps(vec,a); } |
| 103 |
|
F32vec4& operator -=(F32vec4 &a) { return *this = _mm_sub_ps(vec,a); } |
| 104 |
|
F32vec4& operator *=(F32vec4 &a) { return *this = _mm_mul_ps(vec,a); } |
| 105 |
|
F32vec4& operator /=(F32vec4 &a) { return *this = _mm_div_ps(vec,a); } |
| 106 |
|
F32vec4& operator &=(F32vec4 &a) { return *this = _mm_and_ps(vec,a); } |
| 107 |
|
F32vec4& operator |=(F32vec4 &a) { return *this = _mm_or_ps(vec,a); } |
| 108 |
|
F32vec4& operator ^=(F32vec4 &a) { return *this = _mm_xor_ps(vec,a); } |
| 109 |
|
|
| 110 |
|
/* Horizontal Add */ |
| 111 |
|
friend float add_horizontal(F32vec4 &a) |
| 112 |
|
{ |
| 113 |
|
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)))); |
| 114 |
|
return ftemp[0]; |
| 115 |
|
} |
| 116 |
|
|
| 117 |
|
/* Square Root */ |
| 118 |
|
friend F32vec4 sqrt(const F32vec4 &a) { return _mm_sqrt_ps(a); } |
| 119 |
|
/* Reciprocal */ |
| 120 |
|
friend F32vec4 rcp(const F32vec4 &a) { return _mm_rcp_ps(a); } |
| 121 |
|
/* Reciprocal Square Root */ |
| 122 |
|
friend F32vec4 rsqrt(const F32vec4 &a) { return _mm_rsqrt_ps(a); } |
| 123 |
|
|
| 124 |
|
/* NewtonRaphson Reciprocal |
| 125 |
|
[2 * rcpps(x) - (x * rcpps(x) * rcpps(x))] */ |
| 126 |
|
friend F32vec4 rcp_nr(const F32vec4 &a) |
| 127 |
|
{ |
| 128 |
|
F32vec4 Ra0 = _mm_rcp_ps(a); |
| 129 |
|
return _mm_sub_ps(_mm_add_ps(Ra0, Ra0), _mm_mul_ps(_mm_mul_ps(Ra0, a), Ra0)); |
| 130 |
|
} |
| 131 |
|
|
| 132 |
|
/* NewtonRaphson Reciprocal Square Root |
| 133 |
|
0.5 * rsqrtps * (3 - x * rsqrtps(x) * rsqrtps(x)) */ |
| 134 |
|
#pragma warning(push) |
| 135 |
|
#pragma warning(disable : 4640) |
| 136 |
|
friend F32vec4 rsqrt_nr(const F32vec4 &a) |
| 137 |
|
{ |
| 138 |
|
static const F32vec4 fvecf0pt5(0.5f); |
| 139 |
|
static const F32vec4 fvecf3pt0(3.0f); |
| 140 |
|
F32vec4 Ra0 = _mm_rsqrt_ps(a); |
| 141 |
|
return (fvecf0pt5 * Ra0) * (fvecf3pt0 - (a * Ra0) * Ra0); |
| 142 |
|
} |
| 143 |
|
#pragma warning(pop) |
| 144 |
|
|
| 145 |
|
/* Compares: Mask is returned */ |
| 146 |
|
/* Macros expand to all compare intrinsics. Example: |
| 147 |
|
friend F32vec4 cmpeq(const F32vec4 &a, const F32vec4 &b) |
| 148 |
|
{ return _mm_cmpeq_ps(a,b);} */ |
| 149 |
|
#define Fvec32s4_COMP(op) \ |
| 150 |
|
friend F32vec4 cmp##op (const F32vec4 &a, const F32vec4 &b) { return _mm_cmp##op##_ps(a,b); } |
| 151 |
|
Fvec32s4_COMP(eq) /* expanded to cmpeq(a,b) */ |
| 152 |
|
Fvec32s4_COMP(lt) /* expanded to cmplt(a,b) */ |
| 153 |
|
Fvec32s4_COMP(le) /* expanded to cmple(a,b) */ |
| 154 |
|
Fvec32s4_COMP(gt) /* expanded to cmpgt(a,b) */ |
| 155 |
|
Fvec32s4_COMP(ge) /* expanded to cmpge(a,b) */ |
| 156 |
|
Fvec32s4_COMP(neq) /* expanded to cmpneq(a,b) */ |
| 157 |
|
Fvec32s4_COMP(nlt) /* expanded to cmpnlt(a,b) */ |
| 158 |
|
Fvec32s4_COMP(nle) /* expanded to cmpnle(a,b) */ |
| 159 |
|
Fvec32s4_COMP(ngt) /* expanded to cmpngt(a,b) */ |
| 160 |
|
Fvec32s4_COMP(nge) /* expanded to cmpnge(a,b) */ |
| 161 |
|
#undef Fvec32s4_COMP |
| 162 |
|
|
| 163 |
|
/* Min and Max */ |
| 164 |
|
friend F32vec4 simd_min(const F32vec4 &a, const F32vec4 &b) { return _mm_min_ps(a,b); } |
| 165 |
|
friend F32vec4 simd_max(const F32vec4 &a, const F32vec4 &b) { return _mm_max_ps(a,b); } |
| 166 |
|
|
| 167 |
|
/* Debug Features */ |
| 168 |
|
#if defined(_ENABLE_VEC_DEBUG) |
| 169 |
|
/* Output */ |
| 170 |
|
friend std::ostream & operator<<(std::ostream & os, const F32vec4 &a) |
| 171 |
|
{ |
| 172 |
|
/* To use: cout << "Elements of F32vec4 fvec are: " << fvec; */ |
| 173 |
|
float *fp = (float*)&a; |
| 174 |
|
os << "[3]:" << *(fp+3) |
| 175 |
|
<< " [2]:" << *(fp+2) |
| 176 |
|
<< " [1]:" << *(fp+1) |
| 177 |
|
<< " [0]:" << *fp; |
| 178 |
|
return os; |
| 179 |
|
} |
| 180 |
|
#endif |
| 181 |
|
/* Element Access Only, no modifications to elements*/ |
| 182 |
|
const float& operator[](int i) const |
| 183 |
|
{ |
| 184 |
|
/* Assert enabled only during debug /DDEBUG */ |
| 185 |
|
assert((0 <= i) && (i <= 3)); /* User should only access elements 0-3 */ |
| 186 |
|
float *fp = (float*)&vec; |
| 187 |
|
return *(fp+i); |
| 188 |
|
} |
| 189 |
|
/* Element Access and Modification*/ |
| 190 |
|
float& operator[](int i) |
| 191 |
|
{ |
| 192 |
|
/* Assert enabled only during debug /DDEBUG */ |
| 193 |
|
assert((0 <= i) && (i <= 3)); /* User should only access elements 0-3 */ |
| 194 |
|
float *fp = (float*)&vec; |
| 195 |
|
return *(fp+i); |
| 196 |
|
} |
| 197 |
|
}; |
| 198 |
|
|
| 199 |
|
/* Miscellaneous */ |
| 200 |
|
|
| 201 |
|
/* Interleave low order data elements of a and b into destination */ |
| 202 |
|
inline F32vec4 unpack_low(const F32vec4 &a, const F32vec4 &b) |
| 203 |
|
{ return _mm_unpacklo_ps(a, b); } |
| 204 |
|
|
| 205 |
|
/* Interleave high order data elements of a and b into target */ |
| 206 |
|
inline F32vec4 unpack_high(const F32vec4 &a, const F32vec4 &b) |
| 207 |
|
{ return _mm_unpackhi_ps(a, b); } |
| 208 |
|
|
| 209 |
|
/* Move Mask to Integer returns 4 bit mask formed of most significant bits of a */ |
| 210 |
|
inline int move_mask(const F32vec4 &a) |
| 211 |
|
{ return _mm_movemask_ps(a);} |
| 212 |
|
|
| 213 |
|
/* Data Motion Functions */ |
| 214 |
|
|
| 215 |
|
/* Load Unaligned loadu_ps: Unaligned */ |
| 216 |
|
inline void loadu(F32vec4 &a, float *p) |
| 217 |
|
{ a = _mm_loadu_ps(p); } |
| 218 |
|
|
| 219 |
|
/* Store Temporal storeu_ps: Unaligned */ |
| 220 |
|
inline void storeu(float *p, const F32vec4 &a) |
| 221 |
|
{ _mm_storeu_ps(p, a); } |
| 222 |
|
|
| 223 |
|
/* Cacheability Support */ |
| 224 |
|
|
| 225 |
|
/* Non-Temporal Store */ |
| 226 |
|
inline void store_nta(float *p, F32vec4 &a) |
| 227 |
|
{ _mm_stream_ps(p,a);} |
| 228 |
|
|
| 229 |
|
/* Conditional Selects:*/ |
| 230 |
|
/*(a OP b)? c : d; where OP is any compare operator |
| 231 |
|
Macros expand to conditional selects which use all compare intrinsics. |
| 232 |
|
Example: |
| 233 |
|
friend F32vec4 select_eq(const F32vec4 &a, const F32vec4 &b, const F32vec4 &c, const F32vec4 &d) |
| 234 |
|
{ |
| 235 |
|
F32vec4 mask = _mm_cmpeq_ps(a,b); |
| 236 |
|
return( (mask & c) | F32vec4((_mm_andnot_ps(mask,d)))); |
| 237 |
|
} |
| 238 |
|
*/ |
| 239 |
|
|
| 240 |
|
#define Fvec32s4_SELECT(op) \ |
| 241 |
|
inline F32vec4 select_##op (const F32vec4 &a, const F32vec4 &b, const F32vec4 &c, const F32vec4 &d) \ |
| 242 |
|
{ \ |
| 243 |
|
F32vec4 mask = _mm_cmp##op##_ps(a,b); \ |
| 244 |
|
return( (mask & c) | F32vec4((_mm_andnot_ps(mask,d)))); \ |
| 245 |
|
} |
| 246 |
|
Fvec32s4_SELECT(eq) /* generates select_eq(a,b) */ |
| 247 |
|
Fvec32s4_SELECT(lt) /* generates select_lt(a,b) */ |
| 248 |
|
Fvec32s4_SELECT(le) /* generates select_le(a,b) */ |
| 249 |
|
Fvec32s4_SELECT(gt) /* generates select_gt(a,b) */ |
| 250 |
|
Fvec32s4_SELECT(ge) /* generates select_ge(a,b) */ |
| 251 |
|
Fvec32s4_SELECT(neq) /* generates select_neq(a,b) */ |
| 252 |
|
Fvec32s4_SELECT(nlt) /* generates select_nlt(a,b) */ |
| 253 |
|
Fvec32s4_SELECT(nle) /* generates select_nle(a,b) */ |
| 254 |
|
Fvec32s4_SELECT(ngt) /* generates select_ngt(a,b) */ |
| 255 |
|
Fvec32s4_SELECT(nge) /* generates select_nge(a,b) */ |
| 256 |
|
#undef Fvec32s4_SELECT |
| 257 |
|
|
| 258 |
|
|
| 259 |
|
/* Streaming SIMD Extensions Integer Intrinsics */ |
| 260 |
|
|
| 261 |
|
/* Max and Min */ |
| 262 |
|
inline Is16vec4 simd_max(const Is16vec4 &a, const Is16vec4 &b) { return _m_pmaxsw(a,b);} |
| 263 |
|
inline Is16vec4 simd_min(const Is16vec4 &a, const Is16vec4 &b) { return _m_pminsw(a,b);} |
| 264 |
|
inline Iu8vec8 simd_max(const Iu8vec8 &a, const Iu8vec8 &b) { return _m_pmaxub(a,b);} |
| 265 |
|
inline Iu8vec8 simd_min(const Iu8vec8 &a, const Iu8vec8 &b) { return _m_pminub(a,b);} |
| 266 |
|
|
| 267 |
|
/* Average */ |
| 268 |
|
inline Iu16vec4 simd_avg(const Iu16vec4 &a, const Iu16vec4 &b) { return _m_pavgw(a,b); } |
| 269 |
|
inline Iu8vec8 simd_avg(const Iu8vec8 &a, const Iu8vec8 &b) { return _m_pavgb(a,b); } |
| 270 |
|
|
| 271 |
|
/* Move ByteMask To Int: returns mask formed from most sig bits of each vec of a */ |
| 272 |
|
inline int move_mask(const I8vec8 &a) { return _m_pmovmskb(a);} |
| 273 |
|
|
| 274 |
|
/* Packed Multiply High Unsigned */ |
| 275 |
|
inline Iu16vec4 mul_high(const Iu16vec4 &a, const Iu16vec4 &b) { return _m_pmulhuw(a,b); } |
| 276 |
|
|
| 277 |
|
/* Byte Mask Write: Write bytes if most significant bit in each corresponding byte is set */ |
| 278 |
|
inline void mask_move(const I8vec8 &a, const I8vec8 &b, char *addr) { _m_maskmovq(a, b, addr); } |
| 279 |
|
|
| 280 |
|
/* Data Motion: Store Non Temporal */ |
| 281 |
|
inline void store_nta(__m64 *p, M64 &a) { _mm_stream_pi(p,a); } |
| 282 |
|
|
| 283 |
|
/* Conversions between ivec <-> fvec */ |
| 284 |
|
|
| 285 |
|
/* Convert first element of F32vec4 to int with truncation */ |
| 286 |
|
inline int F32vec4ToInt(const F32vec4 &a) |
| 297 |
|
__m64 result; |
| 298 |
|
result = _mm_cvtt_ps2pi(a); |
| 299 |
|
return Is32vec2(result); |
| 300 |
|
|
| 301 |
|
} |
| 302 |
|
|
| 303 |
|
/* Convert the 32-bit int i to an SP FP value; the upper three SP FP values are passed through from a. */ |
| 304 |
|
inline F32vec4 IntToF32vec4(const F32vec4 &a, int i) |
| 305 |
|
{ |
| 306 |
|
|
| 307 |
|
__m128 result; |
| 308 |
|
result = _mm_cvt_si2ss(a,i); |
| 309 |
|
return F32vec4(result); |
| 310 |
|
|
| 311 |
|
} |
| 312 |
|
|
| 313 |
|
/* Convert the two 32-bit integer values in b to two SP FP values; the upper two SP FP values are passed from a. */ |
| 314 |
|
inline F32vec4 Is32vec2ToF32vec4(const F32vec4 &a, const Is32vec2 &b) |
| 315 |
|
{ |
| 316 |
|
|
| 317 |
|
__m128 result; |
| 318 |
|
result = _mm_cvt_pi2ps(a,b); |
| 319 |
|
return F32vec4(result); |
| 320 |
|
} |
| 321 |
|
|
| 322 |
|
class F32vec1 |
| 323 |
|
{ |
| 324 |
|
protected: |
| 325 |
|
__m128 vec; |
| 326 |
|
public: |
| 327 |
|
|
| 328 |
|
/* Constructors: 1 float */ |
| 329 |
|
F32vec1() {} |
| 330 |
|
|
| 331 |
|
F32vec1(int i) { vec = _mm_cvt_si2ss(vec,i);}; |
| 332 |
|
|
| 333 |
|
/* Initialize each of 4 SP FPs with same float */ |
| 334 |
|
EXPLICIT F32vec1(float f) { vec = _mm_set_ss(f); } |
| 335 |
|
|
| 336 |
|
/* Initialize each of 4 SP FPs with same float */ |
| 337 |
|
EXPLICIT F32vec1(double d) { vec = _mm_set_ss((float) d); } |
| 338 |
|
|
| 339 |
|
/* initialize with __m128 data type */ |
| 340 |
|
F32vec1(__m128 m) { vec = m; } |
| 341 |
|
|
| 342 |
|
/* Conversion functions */ |
| 343 |
|
operator __m128() const { return vec; } /* Convert to float */ |
| 344 |
|
|
| 345 |
|
/* Logical Operators */ |
| 346 |
|
friend F32vec1 operator &(const F32vec1 &a, const F32vec1 &b) { return _mm_and_ps(a,b); } |
| 347 |
|
friend F32vec1 operator |(const F32vec1 &a, const F32vec1 &b) { return _mm_or_ps(a,b); } |
| 348 |
|
friend F32vec1 operator ^(const F32vec1 &a, const F32vec1 &b) { return _mm_xor_ps(a,b); } |
| 349 |
|
|
| 350 |
|
/* Arithmetic Operators */ |
| 351 |
|
friend F32vec1 operator +(const F32vec1 &a, const F32vec1 &b) { return _mm_add_ss(a,b); } |
| 352 |
|
friend F32vec1 operator -(const F32vec1 &a, const F32vec1 &b) { return _mm_sub_ss(a,b); } |
| 353 |
|
friend F32vec1 operator *(const F32vec1 &a, const F32vec1 &b) { return _mm_mul_ss(a,b); } |
| 354 |
|
friend F32vec1 operator /(const F32vec1 &a, const F32vec1 &b) { return _mm_div_ss(a,b); } |
| 355 |
|
|
| 356 |
|
F32vec1& operator +=(F32vec1 &a) { return *this = _mm_add_ss(vec,a); } |
| 357 |
|
F32vec1& operator -=(F32vec1 &a) { return *this = _mm_sub_ss(vec,a); } |
| 358 |
|
F32vec1& operator *=(F32vec1 &a) { return *this = _mm_mul_ss(vec,a); } |
| 359 |
|
F32vec1& operator /=(F32vec1 &a) { return *this = _mm_div_ss(vec,a); } |
| 360 |
|
F32vec1& operator &=(F32vec1 &a) { return *this = _mm_and_ps(vec,a); } |
| 361 |
|
F32vec1& operator |=(F32vec1 &a) { return *this = _mm_or_ps(vec,a); } |
| 362 |
|
F32vec1& operator ^=(F32vec1 &a) { return *this = _mm_xor_ps(vec,a); } |
| 363 |
|
|
| 364 |
|
|
| 365 |
|
/* Square Root */ |
| 366 |
|
friend F32vec1 sqrt(const F32vec1 &a) { return _mm_sqrt_ss(a); } |
| 367 |
|
/* Reciprocal */ |
| 368 |
|
friend F32vec1 rcp(const F32vec1 &a) { return _mm_rcp_ss(a); } |
| 369 |
|
/* Reciprocal Square Root */ |
| 370 |
|
friend F32vec1 rsqrt(const F32vec1 &a) { return _mm_rsqrt_ss(a); } |
| 371 |
|
|
| 372 |
|
/* NewtonRaphson Reciprocal |
| 373 |
|
[2 * rcpss(x) - (x * rcpss(x) * rcpss(x))] */ |
| 374 |
|
friend F32vec1 rcp_nr(const F32vec1 &a) |
| 375 |
|
{ |
| 376 |
|
F32vec1 Ra0 = _mm_rcp_ss(a); |
| 377 |
|
return _mm_sub_ss(_mm_add_ss(Ra0, Ra0), _mm_mul_ss(_mm_mul_ss(Ra0, a), Ra0)); |
| 378 |
|
} |
| 379 |
|
|
| 380 |
|
/* NewtonRaphson Reciprocal Square Root |
| 381 |
|
0.5 * rsqrtss * (3 - x * rsqrtss(x) * rsqrtss(x)) */ |
| 382 |
|
#pragma warning(push) |
| 383 |
|
#pragma warning(disable : 4640) |
| 384 |
|
friend F32vec1 rsqrt_nr(const F32vec1 &a) |
| 385 |
|
{ |
| 386 |
|
static const F32vec1 fvecf0pt5(0.5f); |
| 387 |
|
static const F32vec1 fvecf3pt0(3.0f); |
| 388 |
|
F32vec1 Ra0 = _mm_rsqrt_ss(a); |
| 389 |
|
return (fvecf0pt5 * Ra0) * (fvecf3pt0 - (a * Ra0) * Ra0); |
| 390 |
|
} |
| 391 |
|
#pragma warning(pop) |
| 392 |
|
|
| 393 |
|
/* Compares: Mask is returned */ |
| 394 |
|
/* Macros expand to all compare intrinsics. Example: |
| 395 |
|
friend F32vec1 cmpeq(const F32vec1 &a, const F32vec1 &b) |
| 396 |
|
{ return _mm_cmpeq_ss(a,b);} */ |
| 397 |
|
#define Fvec32s1_COMP(op) \ |
| 398 |
|
friend F32vec1 cmp##op (const F32vec1 &a, const F32vec1 &b) { return _mm_cmp##op##_ss(a,b); } |
| 399 |
|
Fvec32s1_COMP(eq) /* expanded to cmpeq(a,b) */ |
| 400 |
|
Fvec32s1_COMP(lt) /* expanded to cmplt(a,b) */ |
| 401 |
|
Fvec32s1_COMP(le) /* expanded to cmple(a,b) */ |
| 402 |
|
Fvec32s1_COMP(gt) /* expanded to cmpgt(a,b) */ |
| 403 |
|
Fvec32s1_COMP(ge) /* expanded to cmpge(a,b) */ |
| 404 |
|
Fvec32s1_COMP(neq) /* expanded to cmpneq(a,b) */ |
| 405 |
|
Fvec32s1_COMP(nlt) /* expanded to cmpnlt(a,b) */ |
| 406 |
|
Fvec32s1_COMP(nle) /* expanded to cmpnle(a,b) */ |
| 407 |
|
Fvec32s1_COMP(ngt) /* expanded to cmpngt(a,b) */ |
| 408 |
|
Fvec32s1_COMP(nge) /* expanded to cmpnge(a,b) */ |
| 409 |
|
#undef Fvec32s1_COMP |
| 410 |
|
|
| 411 |
|
/* Min and Max */ |
| 412 |
|
friend F32vec1 simd_min(const F32vec1 &a, const F32vec1 &b) { return _mm_min_ss(a,b); } |
| 413 |
|
friend F32vec1 simd_max(const F32vec1 &a, const F32vec1 &b) { return _mm_max_ss(a,b); } |
| 414 |
|
|
| 415 |
|
/* Debug Features */ |
| 416 |
|
#if defined(_ENABLE_VEC_DEBUG) |
| 417 |
|
/* Output */ |
| 418 |
|
friend std::ostream & operator<<(std::ostream & os, const F32vec1 &a) |
| 419 |
|
{ |
| 420 |
|
/* To use: cout << "Elements of F32vec1 fvec are: " << fvec; */ |
| 421 |
|
float *fp = (float*)&a; |
| 422 |
|
os << "float:" << *fp; |
| 423 |
|
return os; |
| 424 |
|
} |
| 425 |
|
#endif |
| 426 |
|
|
| 427 |
|
}; |
| 428 |
|
|
| 429 |
|
/* Conditional Selects:*/ |
| 430 |
|
/*(a OP b)? c : d; where OP is any compare operator |
| 431 |
|
Macros expand to conditional selects which use all compare intrinsics. |
| 432 |
|
Example: |
| 433 |
|
friend F32vec1 select_eq(const F32vec1 &a, const F32vec1 &b, const F32vec1 &c, const F32vec1 &d) |
| 434 |
|
{ |
| 435 |
|
F32vec1 mask = _mm_cmpeq_ss(a,b); |
| 436 |
|
return( (mask & c) | F32vec1((_mm_andnot_ps(mask,d)))); |
| 437 |
|
} |
| 438 |
|
*/ |
| 439 |
|
|
| 440 |
|
#define Fvec32s1_SELECT(op) \ |
| 441 |
|
inline F32vec1 select_##op (const F32vec1 &a, const F32vec1 &b, const F32vec1 &c, const F32vec1 &d) \ |
| 442 |
|
{ \ |
| 443 |
|
F32vec1 mask = _mm_cmp##op##_ss(a,b); \ |
| 444 |
|
return( (mask & c) | F32vec1((_mm_andnot_ps(mask,d)))); \ |
| 445 |
|
} |
| 446 |
|
Fvec32s1_SELECT(eq) /* generates select_eq(a,b) */ |
| 447 |
|
Fvec32s1_SELECT(lt) /* generates select_lt(a,b) */ |
| 448 |
|
Fvec32s1_SELECT(le) /* generates select_le(a,b) */ |
| 449 |
|
Fvec32s1_SELECT(gt) /* generates select_gt(a,b) */ |
| 450 |
|
Fvec32s1_SELECT(ge) /* generates select_ge(a,b) */ |
| 451 |
|
Fvec32s1_SELECT(neq) /* generates select_neq(a,b) */ |
| 452 |
|
Fvec32s1_SELECT(nlt) /* generates select_nlt(a,b) */ |
| 453 |
|
Fvec32s1_SELECT(nle) /* generates select_nle(a,b) */ |
| 454 |
|
Fvec32s1_SELECT(ngt) /* generates select_ngt(a,b) */ |
| 455 |
|
Fvec32s1_SELECT(nge) /* generates select_nge(a,b) */ |
| 456 |
|
#undef Fvec32s1_SELECT |
| 457 |
|
|
| 458 |
|
/* Conversions between ivec <-> fvec */ |
| 459 |
|
|
| 460 |
|
/* Convert F32vec1 to int */ |
| 461 |
|
inline int F32vec1ToInt(const F32vec1 &a) |
| 462 |
|
{ |
| 463 |
|
return _mm_cvtt_ss2si(a); |
| 464 |
|
} |
| 465 |
|
|
| 466 |
|
|
| 467 |
|
|
| 468 |
|
#pragma pack(pop) /* 16-B aligned */ |
| 469 |
|
|
| 470 |
|
#ifdef _MSC_VER |
| 471 |
|
#pragma pack(pop) |
| 472 |
|
#endif /* _MSC_VER */ |
| 473 |
|
|
| 474 |
|
#endif /* defined(_M_CEE_PURE) */ |
| 475 |
|
|
| 476 |
|
#endif /* RC_INVOKED */ |
| 477 |
|
#endif /* _FVEC_H_INCLUDED */ |
| 478 |
|
|
| 479 |
|
|
| 480 |
|
|
|
|
|