/** *** Copyright (C) 1985-1999 Intel Corporation. All rights reserved. *** *** The information and source code contained herein is the exclusive *** property of Intel Corporation and may not be disclosed, examined *** or reproduced in whole or in part without explicit written authorization *** from the company. *** **/ /* * Definition of a C++ class interface to Streaming SIMD Extension intrinsics. * * * File name : fvec.h Fvec class definitions * * Concept: A C++ abstraction of Streaming SIMD Extensions designed to improve * * programmer productivity. Speed and accuracy are sacrificed for utility. * * Facilitates an easy transition to compiler intrinsics * * or assembly language. * * F32vec4: 4 packed single precision * 32-bit floating point numbers */ #ifndef FVEC_H_INCLUDED #define FVEC_H_INCLUDED #if !defined __cplusplus #error ERROR: This file is only supported in C++ compilations! #endif /* !__cplusplus */ #include /* Streaming SIMD Extensions Intrinsics include file */ #include #include /* Define _ENABLE_VEC_DEBUG to enable std::ostream inserters for debug output */ #if defined(_ENABLE_VEC_DEBUG) #include #endif #pragma pack(push,16) /* Must ensure class & union 16-B aligned */ /* If using MSVC5.0, explicit keyword should be used */ #if (_MSC_VER >= 1100) #define EXPLICIT explicit #else #if (__ICL) #define EXPLICIT __explicit /* If MSVC4.x & ICL, use __explicit */ #else #define EXPLICIT /* nothing */ #pragma message( "explicit keyword not recognized") #endif #endif class F32vec4 { protected: __m128 vec; public: /* Constructors: __m128, 4 floats, 1 float */ F32vec4() {} /* initialize 4 SP FP with __m128 data type */ F32vec4(__m128 m) { vec = m;} /* initialize 4 SP FPs with 4 floats */ F32vec4(float f3, float f2, float f1, float f0) { vec= _mm_set_ps(f3,f2,f1,f0); } /* Explicitly initialize each of 4 SP FPs with same float */ EXPLICIT F32vec4(float f) { vec = _mm_set_ps1(f); } /* Explicitly initialize each of 4 SP FPs with same double */ EXPLICIT F32vec4(double d) { vec = _mm_set_ps1((float) d); } /* Assignment operations */ F32vec4& operator =(float f) { vec = _mm_set_ps1(f); return *this; } F32vec4& operator =(double d) { vec = _mm_set_ps1((float) d); return *this; } /* Conversion functions */ operator __m128() const { return vec; } /* Convert to __m128 */ /* Logical Operators */ friend F32vec4 operator &(const F32vec4 &a, const F32vec4 &b) { return _mm_and_ps(a,b); } friend F32vec4 operator |(const F32vec4 &a, const F32vec4 &b) { return _mm_or_ps(a,b); } friend F32vec4 operator ^(const F32vec4 &a, const F32vec4 &b) { return _mm_xor_ps(a,b); } /* Arithmetic Operators */ friend F32vec4 operator +(const F32vec4 &a, const F32vec4 &b) { return _mm_add_ps(a,b); } friend F32vec4 operator -(const F32vec4 &a, const F32vec4 &b) { return _mm_sub_ps(a,b); } friend F32vec4 operator *(const F32vec4 &a, const F32vec4 &b) { return _mm_mul_ps(a,b); } friend F32vec4 operator /(const F32vec4 &a, const F32vec4 &b) { return _mm_div_ps(a,b); } F32vec4& operator =(const F32vec4 &a) { vec = a.vec; return *this; } F32vec4& operator =(const __m128 &avec) { vec = avec; return *this; } F32vec4& operator +=(F32vec4 &a) { return *this = _mm_add_ps(vec,a); } F32vec4& operator -=(F32vec4 &a) { return *this = _mm_sub_ps(vec,a); } F32vec4& operator *=(F32vec4 &a) { return *this = _mm_mul_ps(vec,a); } F32vec4& operator /=(F32vec4 &a) { return *this = _mm_div_ps(vec,a); } F32vec4& operator &=(F32vec4 &a) { return *this = _mm_and_ps(vec,a); } F32vec4& operator |=(F32vec4 &a) { return *this = _mm_or_ps(vec,a); } F32vec4& operator ^=(F32vec4 &a) { return *this = _mm_xor_ps(vec,a); } /* Horizontal Add */ friend float add_horizontal(F32vec4 &a) { 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)))); return ftemp[0]; } /* Square Root */ friend F32vec4 sqrt(const F32vec4 &a) { return _mm_sqrt_ps(a); } /* Reciprocal */ friend F32vec4 rcp(const F32vec4 &a) { return _mm_rcp_ps(a); } /* Reciprocal Square Root */ friend F32vec4 rsqrt(const F32vec4 &a) { return _mm_rsqrt_ps(a); } /* NewtonRaphson Reciprocal [2 * rcpps(x) - (x * rcpps(x) * rcpps(x))] */ friend F32vec4 rcp_nr(const F32vec4 &a) { F32vec4 Ra0 = _mm_rcp_ps(a); return _mm_sub_ps(_mm_add_ps(Ra0, Ra0), _mm_mul_ps(_mm_mul_ps(Ra0, a), Ra0)); } /* NewtonRaphson Reciprocal Square Root 0.5 * rsqrtps * (3 - x * rsqrtps(x) * rsqrtps(x)) */ friend F32vec4 rsqrt_nr(const F32vec4 &a) { static const F32vec4 fvecf0pt5(0.5f); static const F32vec4 fvecf3pt0(3.0f); F32vec4 Ra0 = _mm_rsqrt_ps(a); return (fvecf0pt5 * Ra0) * (fvecf3pt0 - (a * Ra0) * Ra0); } /* Compares: Mask is returned */ /* Macros expand to all compare intrinsics. Example: friend F32vec4 cmpeq(const F32vec4 &a, const F32vec4 &b) { return _mm_cmpeq_ps(a,b);} */ #define Fvec32s4_COMP(op) \ friend F32vec4 cmp##op (const F32vec4 &a, const F32vec4 &b) { return _mm_cmp##op##_ps(a,b); } Fvec32s4_COMP(eq) // expanded to cmpeq(a,b) Fvec32s4_COMP(lt) // expanded to cmplt(a,b) Fvec32s4_COMP(le) // expanded to cmple(a,b) Fvec32s4_COMP(gt) // expanded to cmpgt(a,b) Fvec32s4_COMP(ge) // expanded to cmpge(a,b) Fvec32s4_COMP(neq) // expanded to cmpneq(a,b) Fvec32s4_COMP(nlt) // expanded to cmpnlt(a,b) Fvec32s4_COMP(nle) // expanded to cmpnle(a,b) Fvec32s4_COMP(ngt) // expanded to cmpngt(a,b) Fvec32s4_COMP(nge) // expanded to cmpnge(a,b) #undef Fvec32s4_COMP /* Min and Max */ friend F32vec4 simd_min(const F32vec4 &a, const F32vec4 &b) { return _mm_min_ps(a,b); } friend F32vec4 simd_max(const F32vec4 &a, const F32vec4 &b) { return _mm_max_ps(a,b); } /* Debug Features */ #if defined(_ENABLE_VEC_DEBUG) /* Output */ friend std::ostream & operator<<(std::ostream & os, const F32vec4 &a) { /* To use: cout << "Elements of F32vec4 fvec are: " << fvec; */ float *fp = (float*)&a; os << "[3]:" << *(fp+3) << " [2]:" << *(fp+2) << " [1]:" << *(fp+1) << " [0]:" << *fp; return os; } #endif /* Element Access Only, no modifications to elements*/ const float& operator[](int i) const { /* Assert enabled only during debug /DDEBUG */ assert((0 <= i) && (i <= 3)); /* User should only access elements 0-3 */ float *fp = (float*)&vec; return *(fp+i); } /* Element Access and Modification*/ float& operator[](int i) { /* Assert enabled only during debug /DDEBUG */ assert((0 <= i) && (i <= 3)); /* User should only access elements 0-3 */ float *fp = (float*)&vec; return *(fp+i); } }; /* Miscellaneous */ /* Interleave low order data elements of a and b into destination */ inline F32vec4 unpack_low(const F32vec4 &a, const F32vec4 &b) { return _mm_unpacklo_ps(a, b); } /* Interleave high order data elements of a and b into target */ inline F32vec4 unpack_high(const F32vec4 &a, const F32vec4 &b) { return _mm_unpackhi_ps(a, b); } /* Move Mask to Integer returns 4 bit mask formed of most significant bits of a */ inline int move_mask(const F32vec4 &a) { return _mm_movemask_ps(a);} /* Data Motion Functions */ /* Load Unaligned loadu_ps: Unaligned */ inline void loadu(F32vec4 &a, float *p) { a = _mm_loadu_ps(p); } /* Store Temporal storeu_ps: Unaligned */ inline void storeu(float *p, const F32vec4 &a) { _mm_storeu_ps(p, a); } /* Cacheability Support */ /* Non-Temporal Store */ inline void store_nta(float *p, F32vec4 &a) { _mm_stream_ps(p,a);} /* Conditional Selects:*/ /*(a OP b)? c : d; where OP is any compare operator Macros expand to conditional selects which use all compare intrinsics. Example: friend F32vec4 select_eq(const F32vec4 &a, const F32vec4 &b, const F32vec4 &c, const F32vec4 &d) { F32vec4 mask = _mm_cmpeq_ps(a,b); return( (mask & c) | F32vec4((_mm_andnot_ps(mask,d)))); } */ #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)))); \ } Fvec32s4_SELECT(eq) // generates select_eq(a,b) Fvec32s4_SELECT(lt) // generates select_lt(a,b) Fvec32s4_SELECT(le) // generates select_le(a,b) Fvec32s4_SELECT(gt) // generates select_gt(a,b) Fvec32s4_SELECT(ge) // generates select_ge(a,b) Fvec32s4_SELECT(neq) // generates select_neq(a,b) Fvec32s4_SELECT(nlt) // generates select_nlt(a,b) Fvec32s4_SELECT(nle) // generates select_nle(a,b) Fvec32s4_SELECT(ngt) // generates select_ngt(a,b) Fvec32s4_SELECT(nge) // generates select_nge(a,b) #undef Fvec32s4_SELECT /* Streaming SIMD Extensions Integer Intrinsics */ /* Max and Min */ inline Is16vec4 simd_max(const Is16vec4 &a, const Is16vec4 &b) { return _m_pmaxsw(a,b);} inline Is16vec4 simd_min(const Is16vec4 &a, const Is16vec4 &b) { return _m_pminsw(a,b);} inline Iu8vec8 simd_max(const Iu8vec8 &a, const Iu8vec8 &b) { return _m_pmaxub(a,b);} inline Iu8vec8 simd_min(const Iu8vec8 &a, const Iu8vec8 &b) { return _m_pminub(a,b);} /* Average */ inline Iu16vec4 simd_avg(const Iu16vec4 &a, const Iu16vec4 &b) { return _m_pavgw(a,b); } inline Iu8vec8 simd_avg(const Iu8vec8 &a, const Iu8vec8 &b) { return _m_pavgb(a,b); } /* Move ByteMask To Int: returns mask formed from most sig bits of each vec of a */ inline int move_mask(const I8vec8 &a) { return _m_pmovmskb(a);} /* Packed Multiply High Unsigned */ inline Iu16vec4 mul_high(const Iu16vec4 &a, const Iu16vec4 &b) { return _m_pmulhuw(a,b); } /* Byte Mask Write: Write bytes if most significant bit in each corresponding byte is set */ inline void mask_move(const I8vec8 &a, const I8vec8 &b, char *addr) { _m_maskmovq(a, b, addr); } /* Data Motion: Store Non Temporal */ inline void store_nta(__m64 *p, M64 &a) { _mm_stream_pi(p,a); } /* Conversions between ivec <-> fvec */ /* Convert first element of F32vec4 to int with truncation */ inline int F32vec4ToInt(const F32vec4 &a) { return _mm_cvtt_ss2si(a); } /* Convert two lower SP FP values of a to Is32vec2 with truncation */ inline Is32vec2 F32vec4ToIs32vec2 (const F32vec4 &a) { __m64 result; result = _mm_cvtt_ps2pi(a); return Is32vec2(result); } /* Convert the 32-bit int i to an SP FP value; the upper three SP FP values are passed through from a. */ inline F32vec4 IntToF32vec4(const F32vec4 &a, int i) { __m128 result; result = _mm_cvt_si2ss(a,i); return F32vec4(result); } /* Convert the two 32-bit integer values in b to two SP FP values; the upper two SP FP values are passed from a. */ inline F32vec4 Is32vec2ToF32vec4(const F32vec4 &a, const Is32vec2 &b) { __m128 result; result = _mm_cvt_pi2ps(a,b); return F32vec4(result); } class F32vec1 { protected: __m128 vec; public: /* Constructors: 1 float */ F32vec1() {} F32vec1(int i) { vec = _mm_cvt_si2ss(vec,i);}; /* Initialize each of 4 SP FPs with same float */ EXPLICIT F32vec1(float f) { vec = _mm_set_ss(f); } /* Initialize each of 4 SP FPs with same float */ EXPLICIT F32vec1(double d) { vec = _mm_set_ss((float) d); } /* initialize with __m128 data type */ F32vec1(__m128 m) { vec = m; } /* Conversion functions */ operator __m128() const { return vec; } /* Convert to float */ /* Logical Operators */ friend F32vec1 operator &(const F32vec1 &a, const F32vec1 &b) { return _mm_and_ps(a,b); } friend F32vec1 operator |(const F32vec1 &a, const F32vec1 &b) { return _mm_or_ps(a,b); } friend F32vec1 operator ^(const F32vec1 &a, const F32vec1 &b) { return _mm_xor_ps(a,b); } /* Arithmetic Operators */ friend F32vec1 operator +(const F32vec1 &a, const F32vec1 &b) { return _mm_add_ss(a,b); } friend F32vec1 operator -(const F32vec1 &a, const F32vec1 &b) { return _mm_sub_ss(a,b); } friend F32vec1 operator *(const F32vec1 &a, const F32vec1 &b) { return _mm_mul_ss(a,b); } friend F32vec1 operator /(const F32vec1 &a, const F32vec1 &b) { return _mm_div_ss(a,b); } F32vec1& operator +=(F32vec1 &a) { return *this = _mm_add_ss(vec,a); } F32vec1& operator -=(F32vec1 &a) { return *this = _mm_sub_ss(vec,a); } F32vec1& operator *=(F32vec1 &a) { return *this = _mm_mul_ss(vec,a); } F32vec1& operator /=(F32vec1 &a) { return *this = _mm_div_ss(vec,a); } F32vec1& operator &=(F32vec1 &a) { return *this = _mm_and_ps(vec,a); } F32vec1& operator |=(F32vec1 &a) { return *this = _mm_or_ps(vec,a); } F32vec1& operator ^=(F32vec1 &a) { return *this = _mm_xor_ps(vec,a); } /* Square Root */ friend F32vec1 sqrt(const F32vec1 &a) { return _mm_sqrt_ss(a); } /* Reciprocal */ friend F32vec1 rcp(const F32vec1 &a) { return _mm_rcp_ss(a); } /* Reciprocal Square Root */ friend F32vec1 rsqrt(const F32vec1 &a) { return _mm_rsqrt_ss(a); } /* NewtonRaphson Reciprocal [2 * rcpss(x) - (x * rcpss(x) * rcpss(x))] */ friend F32vec1 rcp_nr(const F32vec1 &a) { F32vec1 Ra0 = _mm_rcp_ss(a); return _mm_sub_ss(_mm_add_ss(Ra0, Ra0), _mm_mul_ss(_mm_mul_ss(Ra0, a), Ra0)); } /* NewtonRaphson Reciprocal Square Root 0.5 * rsqrtss * (3 - x * rsqrtss(x) * rsqrtss(x)) */ friend F32vec1 rsqrt_nr(const F32vec1 &a) { static const F32vec1 fvecf0pt5(0.5f); static const F32vec1 fvecf3pt0(3.0f); F32vec1 Ra0 = _mm_rsqrt_ss(a); return (fvecf0pt5 * Ra0) * (fvecf3pt0 - (a * Ra0) * Ra0); } /* Compares: Mask is returned */ /* Macros expand to all compare intrinsics. Example: friend F32vec1 cmpeq(const F32vec1 &a, const F32vec1 &b) { return _mm_cmpeq_ss(a,b);} */ #define Fvec32s1_COMP(op) \ friend F32vec1 cmp##op (const F32vec1 &a, const F32vec1 &b) { return _mm_cmp##op##_ss(a,b); } Fvec32s1_COMP(eq) // expanded to cmpeq(a,b) Fvec32s1_COMP(lt) // expanded to cmplt(a,b) Fvec32s1_COMP(le) // expanded to cmple(a,b) Fvec32s1_COMP(gt) // expanded to cmpgt(a,b) Fvec32s1_COMP(ge) // expanded to cmpge(a,b) Fvec32s1_COMP(neq) // expanded to cmpneq(a,b) Fvec32s1_COMP(nlt) // expanded to cmpnlt(a,b) Fvec32s1_COMP(nle) // expanded to cmpnle(a,b) Fvec32s1_COMP(ngt) // expanded to cmpngt(a,b) Fvec32s1_COMP(nge) // expanded to cmpnge(a,b) #undef Fvec32s1_COMP /* Min and Max */ friend F32vec1 simd_min(const F32vec1 &a, const F32vec1 &b) { return _mm_min_ss(a,b); } friend F32vec1 simd_max(const F32vec1 &a, const F32vec1 &b) { return _mm_max_ss(a,b); } /* Debug Features */ #if defined(_ENABLE_VEC_DEBUG) /* Output */ friend std::ostream & operator<<(std::ostream & os, const F32vec1 &a) { /* To use: cout << "Elements of F32vec1 fvec are: " << fvec; */ float *fp = (float*)&a; os << "float:" << *fp; return os; } #endif }; /* Conditional Selects:*/ /*(a OP b)? c : d; where OP is any compare operator Macros expand to conditional selects which use all compare intrinsics. Example: friend F32vec1 select_eq(const F32vec1 &a, const F32vec1 &b, const F32vec1 &c, const F32vec1 &d) { F32vec1 mask = _mm_cmpeq_ss(a,b); return( (mask & c) | F32vec1((_mm_andnot_ps(mask,d)))); } */ #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)))); \ } Fvec32s1_SELECT(eq) // generates select_eq(a,b) Fvec32s1_SELECT(lt) // generates select_lt(a,b) Fvec32s1_SELECT(le) // generates select_le(a,b) Fvec32s1_SELECT(gt) // generates select_gt(a,b) Fvec32s1_SELECT(ge) // generates select_ge(a,b) Fvec32s1_SELECT(neq) // generates select_neq(a,b) Fvec32s1_SELECT(nlt) // generates select_nlt(a,b) Fvec32s1_SELECT(nle) // generates select_nle(a,b) Fvec32s1_SELECT(ngt) // generates select_ngt(a,b) Fvec32s1_SELECT(nge) // generates select_nge(a,b) #undef Fvec32s1_SELECT /* Conversions between ivec <-> fvec */ /* Convert F32vec1 to int */ inline int F32vec1ToInt(const F32vec1 &a) { return _mm_cvtt_ss2si(a); } #pragma pack(pop) /* 16-B aligned */ #endif /* FVEC_H_INCLUDED */