diff options
Diffstat (limited to 'Eigen/src/Core/arch/SSE/PacketMath.h')
| -rwxr-xr-x[-rw-r--r--] | Eigen/src/Core/arch/SSE/PacketMath.h | 531 |
1 files changed, 349 insertions, 182 deletions
diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 32caaa26c..3832de147 100644..100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -22,9 +22,40 @@ namespace internal { #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) #endif +#ifdef __FMA__ +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 +#endif +#endif + +#if (defined EIGEN_VECTORIZE_AVX) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_MINGW) && (__GXX_ABI_VERSION < 1004) +// With GCC's default ABI version, a __m128 or __m256 are the same types and therefore we cannot +// have overloads for both types without linking error. +// One solution is to increase ABI version using -fabi-version=4 (or greater). +// Otherwise, we workaround this inconvenience by wrapping 128bit types into the following helper +// structure: +template<typename T> +struct eigen_packet_wrapper +{ + EIGEN_ALWAYS_INLINE operator T&() { return m_val; } + EIGEN_ALWAYS_INLINE operator const T&() const { return m_val; } + EIGEN_ALWAYS_INLINE eigen_packet_wrapper() {} + EIGEN_ALWAYS_INLINE eigen_packet_wrapper(const T &v) : m_val(v) {} + EIGEN_ALWAYS_INLINE eigen_packet_wrapper& operator=(const T &v) { + m_val = v; + return *this; + } + + T m_val; +}; +typedef eigen_packet_wrapper<__m128> Packet4f; +typedef eigen_packet_wrapper<__m128i> Packet4i; +typedef eigen_packet_wrapper<__m128d> Packet2d; +#else typedef __m128 Packet4f; typedef __m128i Packet4i; typedef __m128d Packet2d; +#endif template<> struct is_arithmetic<__m128> { enum { value = true }; }; template<> struct is_arithmetic<__m128i> { enum { value = true }; }; @@ -58,51 +89,85 @@ template<> struct is_arithmetic<__m128d> { enum { value = true }; }; const Packet4i p4i_##NAME = pset1<Packet4i>(X) +// Use the packet_traits defined in AVX/PacketMath.h instead if we're going +// to leverage AVX instructions. +#ifndef EIGEN_VECTORIZE_AVX template<> struct packet_traits<float> : default_packet_traits { typedef Packet4f type; + typedef Packet4f half; enum { Vectorizable = 1, AlignedOnScalar = 1, size=4, + HasHalfPacket = 0, HasDiv = 1, HasSin = EIGEN_FAST_MATH, HasCos = EIGEN_FAST_MATH, HasLog = 1, HasExp = 1, - HasSqrt = 1 + HasSqrt = 1, + HasRsqrt = 1, + HasTanh = EIGEN_FAST_MATH, + HasBlend = 1 + +#ifdef EIGEN_VECTORIZE_SSE4_1 + , + HasRound = 1, + HasFloor = 1, + HasCeil = 1 +#endif }; }; template<> struct packet_traits<double> : default_packet_traits { typedef Packet2d type; + typedef Packet2d half; enum { Vectorizable = 1, AlignedOnScalar = 1, size=2, + HasHalfPacket = 0, HasDiv = 1, HasExp = 1, - HasSqrt = 1 + HasSqrt = 1, + HasRsqrt = 1, + HasBlend = 1 + +#ifdef EIGEN_VECTORIZE_SSE4_1 + , + HasRound = 1, + HasFloor = 1, + HasCeil = 1 +#endif }; }; +#endif template<> struct packet_traits<int> : default_packet_traits { typedef Packet4i type; + typedef Packet4i half; enum { - // FIXME check the Has* Vectorizable = 1, AlignedOnScalar = 1, - size=4 + size=4, + + HasBlend = 1 }; }; -template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; }; -template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2}; }; -template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; }; +template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; }; +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; }; +template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; }; -#if defined(_MSC_VER) && (_MSC_VER==1500) +#ifndef EIGEN_VECTORIZE_AVX +template<> struct scalar_div_cost<float,true> { enum { value = 7 }; }; +template<> struct scalar_div_cost<double,true> { enum { value = 8 }; }; +#endif + +#if EIGEN_COMP_MSVC==1500 // Workaround MSVC 9 internal compiler error. // TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode // TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)). @@ -110,14 +175,25 @@ template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { re template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); } template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set_epi32(from,from,from,from); } #else -template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set1_ps(from); } +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps1(from); } template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); } template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set1_epi32(from); } #endif -template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); } -template<> EIGEN_STRONG_INLINE Packet2d plset<double>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); } -template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); } +// GCC generates a shufps instruction for _mm_set1_ps/_mm_load1_ps instead of the more efficient pshufd instruction. +// However, using inrinsics for pset1 makes gcc to generate crappy code in some cases (see bug 203) +// Using inline assembly is also not an option because then gcc fails to reorder properly the instructions. +// Therefore, we introduced the pload1 functions to be used in product kernels for which bug 203 does not apply. +// Also note that with AVX, we want it to generate a vbroadcastss. +#if EIGEN_COMP_GNUC_STRICT && (!defined __AVX__) +template<> EIGEN_STRONG_INLINE Packet4f pload1<Packet4f>(const float *from) { + return vec4f_swizzle1(_mm_load_ss(from),0,0,0,0); +} +#endif + +template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); } +template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); } +template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); } template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_add_pd(a,b); } @@ -139,7 +215,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) } template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { - return psub(_mm_setr_epi32(0,0,0,0), a); + return psub(Packet4i(_mm_setr_epi32(0,0,0,0)), a); } template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; } @@ -166,13 +242,13 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); } -template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) -{ eigen_assert(false && "packet integer division are not supported by SSE"); - return pset1<Packet4i>(0); -} // for some weird raisons, it has to be overloaded for packet of integers template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); } +#ifdef __FMA__ +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); } +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); } +#endif template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_min_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_min_pd(a,b); } @@ -200,6 +276,17 @@ template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const #endif } +#ifdef EIGEN_VECTORIZE_SSE4_1 +template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); } +template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); } + +template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); } +template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); } + +template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); } +template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); } +#endif + template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); } template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); } @@ -216,42 +303,16 @@ template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, con template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(a,b); } template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(a,b); } -template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { - EIGEN_DEBUG_ALIGNED_LOAD -#ifdef EIGEN_ANDROID_SSE_WR -// Workaround for X86 on Android crash on aligned operation. - return _mm_loadu_ps(from); -#else - return _mm_load_ps(from); -#endif - } -template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { - EIGEN_DEBUG_ALIGNED_LOAD -#ifdef EIGEN_ANDROID_SSE_WR -// Workaround for X86 on Android crash on aligned operation. - return _mm_loadu_pd(from); -#else - return _mm_load_pd(from); -#endif - } -template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { - EIGEN_DEBUG_ALIGNED_LOAD -#ifdef EIGEN_ANDROID_SSE_WR -// Workaround for X86 on Android crash on aligned operation. - return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); -#else - return _mm_load_si128(reinterpret_cast<const Packet4i*>(from)); -#endif - } +template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); } +template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); } +template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const __m128i*>(from)); } -#if defined(_MSC_VER) +#if EIGEN_COMP_MSVC template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD - #if (_MSC_VER==1600) + #if (EIGEN_COMP_MSVC==1600) // NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps // (i.e., it does not generate an unaligned load!! - // TODO On most architectures this version should also be faster than a single _mm_loadu_ps - // so we could also enable it for MSVC08 but first we have to make this later does not generate crap when doing so... __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from)); res = _mm_loadh_pi(res, (const __m64*)(from+2)); return res; @@ -259,78 +320,27 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { return _mm_loadu_ps(from); #endif } - template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_pd(from); } - template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); } #else -// Fast unaligned loads. Note that here we cannot directly use intrinsics: this would -// require pointer casting to incompatible pointer types and leads to invalid code -// because of the strict aliasing rule. The "dummy" stuff are required to enforce -// a correct instruction dependency. -// TODO: do the same for MSVC (ICC is compatible) // NOTE: with the code below, MSVC's compiler crashes! -#if defined(__GNUC__) && defined(__i386__) - // bug 195: gcc/i386 emits weird x87 fldl/fstpl instructions for _mm_load_sd - #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1 -#elif defined(__clang__) - // bug 201: Segfaults in __mm_loadh_pd with clang 2.8 - #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1 -#else - #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 0 -#endif - template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD -#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS return _mm_loadu_ps(from); -#else - __m128d res; -#ifdef EIGEN_ANDROID_SSE_WR -// Workaround for X86 on Android crash on aligned operation. - res = _mm_loadu_sd((const double*)(from)) ; -#else - res = _mm_load_sd((const double*)(from)) ; -#endif - res = _mm_loadh_pd(res, (const double*)(from+2)) ; - return _mm_castpd_ps(res); -#endif } +#endif + template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD -#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS return _mm_loadu_pd(from); -#else - __m128d res; -#ifdef EIGEN_ANDROID_SSE_WR -// Workaround for X86 on Android crash on aligned operation. - res = _mm_loadu_sd(from) ; -#else - res = _mm_load_sd(from) ; -#endif - res = _mm_loadh_pd(res,from+1); - return res; -#endif } template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD -#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS - return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); -#else - __m128d res; -#ifdef EIGEN_ANDROID_SSE_WR -// Workaround for X86 on Android crash on aligned operation. - res = _mm_loadu_sd((const double*)(from)) ; -#else - res = _mm_load_sd((const double*)(from)) ; -#endif - res = _mm_loadh_pd(res, (const double*)(from+2)) ; - return _mm_castpd_si128(res); -#endif + return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from)); } -#endif + template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) { @@ -341,70 +351,77 @@ template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) { Packet4i tmp; - tmp = _mm_loadl_epi64(reinterpret_cast<const Packet4i*>(from)); + tmp = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(from)); return vec4i_swizzle1(tmp, 0, 0, 1, 1); } -template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { - EIGEN_DEBUG_ALIGNED_STORE -#ifdef EIGEN_ANDROID_SSE_WR -// Workaround for X86 on Android crash on aligned operation. - _mm_storeu_ps(to, from); -#else - _mm_store_ps(to, from); -#endif - } -template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { - EIGEN_DEBUG_ALIGNED_STORE -#ifdef EIGEN_ANDROID_SSE_WR -// Workaround for X86 on Android crash on aligned operation. - _mm_storeu_pd(to, from); -#else - _mm_store_pd(to, from); -#endif - } -template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { - EIGEN_DEBUG_ALIGNED_STORE -#ifdef EIGEN_ANDROID_SSE_WR -// Workaround for X86 on Android crash on aligned operation. - _mm_storeu_si128(reinterpret_cast<Packet4i*>(to), from); -#else - _mm_store_si128(reinterpret_cast<Packet4i*>(to), from); -#endif - } +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to), from); } + +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from); } -template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { - EIGEN_DEBUG_UNALIGNED_STORE - _mm_storel_pd((to), from); - _mm_storeh_pd((to+1), from); +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) +{ + return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); +} +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride) +{ + return _mm_set_pd(from[1*stride], from[0*stride]); +} +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride) +{ + return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); + } + +template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) +{ + to[stride*0] = _mm_cvtss_f32(from); + to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1)); + to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2)); + to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3)); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride) +{ + to[stride*0] = _mm_cvtsd_f64(from); + to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1)); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride) +{ + to[stride*0] = _mm_cvtsi128_si32(from); + to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1)); + to[stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2)); + to[stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3)); } -template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castps_pd(from)); } -template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castsi128_pd(from)); } // some compilers might be tempted to perform multiple moves instead of using a vector path. template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a) { Packet4f pa = _mm_set_ss(a); - pstore(to, vec4f_swizzle1(pa,0,0,0,0)); + pstore(to, Packet4f(vec4f_swizzle1(pa,0,0,0,0))); } // some compilers might be tempted to perform multiple moves instead of using a vector path. template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double& a) { Packet2d pa = _mm_set_sd(a); - pstore(to, vec2d_swizzle1(pa,0,0)); + pstore(to, Packet2d(vec2d_swizzle1(pa,0,0))); } +#ifndef EIGEN_VECTORIZE_AVX template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +#endif -#if defined(_MSC_VER) && defined(_WIN64) && !defined(__INTEL_COMPILER) +#if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64 // The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 // Direct of the struct members fixed bug #62. template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; } template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; } template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; } -#elif defined(_MSC_VER) && !defined(__INTEL_COMPILER) +#elif EIGEN_COMP_MSVC_STRICT // The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; } template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; } @@ -422,7 +439,6 @@ template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return _mm_shuffle_epi32(a,0x1B); } - template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF)); @@ -443,6 +459,38 @@ template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) #endif } +// with AVX, the default implementations based on pload1 are faster +#ifndef __AVX__ +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet4f>(const float *a, + Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) +{ + a3 = pload<Packet4f>(a); + a0 = vec4f_swizzle1(a3, 0,0,0,0); + a1 = vec4f_swizzle1(a3, 1,1,1,1); + a2 = vec4f_swizzle1(a3, 2,2,2,2); + a3 = vec4f_swizzle1(a3, 3,3,3,3); +} +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet2d>(const double *a, + Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3) +{ +#ifdef EIGEN_VECTORIZE_SSE3 + a0 = _mm_loaddup_pd(a+0); + a1 = _mm_loaddup_pd(a+1); + a2 = _mm_loaddup_pd(a+2); + a3 = _mm_loaddup_pd(a+3); +#else + a1 = pload<Packet2d>(a); + a0 = vec2d_swizzle1(a1, 0,0); + a1 = vec2d_swizzle1(a1, 1,1); + a3 = pload<Packet2d>(a+2); + a2 = vec2d_swizzle1(a3, 0,0); + a3 = vec2d_swizzle1(a3, 1,1); +#endif +} +#endif + EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs) { vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55)); @@ -452,47 +500,17 @@ EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs) } #ifdef EIGEN_VECTORIZE_SSE3 -// TODO implement SSE2 versions as well as integer versions template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) { return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3])); } + template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) { return _mm_hadd_pd(vecs[0], vecs[1]); } -// SSSE3 version: -// EIGEN_STRONG_INLINE Packet4i preduxp(const Packet4i* vecs) -// { -// return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3])); -// } - -template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) -{ - Packet4f tmp0 = _mm_hadd_ps(a,a); - return pfirst(_mm_hadd_ps(tmp0, tmp0)); -} - -template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return pfirst(_mm_hadd_pd(a, a)); } -// SSSE3 version: -// EIGEN_STRONG_INLINE float predux(const Packet4i& a) -// { -// Packet4i tmp0 = _mm_hadd_epi32(a,a); -// return pfirst(_mm_hadd_epi32(tmp0, tmp0)); -// } #else -// SSE2 versions -template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) -{ - Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a)); - return pfirst(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); -} -template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) -{ - return pfirst(_mm_add_sd(a, _mm_unpackhi_pd(a,a))); -} - template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) { Packet4f tmp0, tmp1, tmp2; @@ -513,10 +531,45 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) } #endif // SSE3 +template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures + // (from Nehalem to Haswell) +// #ifdef EIGEN_VECTORIZE_SSE3 +// Packet4f tmp = _mm_add_ps(a, vec4f_swizzle1(a,2,3,2,3)); +// return pfirst<Packet4f>(_mm_hadd_ps(tmp, tmp)); +// #else + Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a)); + return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); +// #endif +} + +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) +{ + // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures + // (from Nehalem to Haswell) +// #ifdef EIGEN_VECTORIZE_SSE3 +// return pfirst<Packet2d>(_mm_hadd_pd(a, a)); +// #else + return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a))); +// #endif +} + +#ifdef EIGEN_VECTORIZE_SSSE3 +template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) +{ + return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3])); +} +template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) +{ + Packet4i tmp0 = _mm_hadd_epi32(a,a); + return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0)); +} +#else template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) { Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a)); - return pfirst(tmp) + pfirst(_mm_shuffle_epi32(tmp, 1)); + return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1)); } template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) @@ -532,18 +585,18 @@ template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) tmp0 = _mm_unpackhi_epi64(tmp0, tmp1); return _mm_add_epi32(tmp0, tmp2); } - +#endif // Other reduction functions: // mul template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) { Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a)); - return pfirst(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); + return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); } template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { - return pfirst(_mm_mul_sd(a, _mm_unpackhi_pd(a,a))); + return pfirst<Packet2d>(_mm_mul_sd(a, _mm_unpackhi_pd(a,a))); } template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) { @@ -559,14 +612,18 @@ template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) { Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a)); - return pfirst(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); + return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); } template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) { - return pfirst(_mm_min_sd(a, _mm_unpackhi_pd(a,a))); + return pfirst<Packet2d>(_mm_min_sd(a, _mm_unpackhi_pd(a,a))); } template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) { +#ifdef EIGEN_VECTORIZE_SSE4_1 + Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); + return pfirst<Packet4i>(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); +#else // after some experiments, it is seems this is the fastest way to implement it // for GCC (eg., it does not like using std::min after the pstore !!) EIGEN_ALIGN16 int aux[4]; @@ -574,20 +631,25 @@ template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) int aux0 = aux[0]<aux[1] ? aux[0] : aux[1]; int aux2 = aux[2]<aux[3] ? aux[2] : aux[3]; return aux0<aux2 ? aux0 : aux2; +#endif // EIGEN_VECTORIZE_SSE4_1 } // max template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) { Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a)); - return pfirst(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); + return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); } template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) { - return pfirst(_mm_max_sd(a, _mm_unpackhi_pd(a,a))); + return pfirst<Packet2d>(_mm_max_sd(a, _mm_unpackhi_pd(a,a))); } template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) { +#ifdef EIGEN_VECTORIZE_SSE4_1 + Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); + return pfirst<Packet4i>(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); +#else // after some experiments, it is seems this is the fastest way to implement it // for GCC (eg., it does not like using std::min after the pstore !!) EIGEN_ALIGN16 int aux[4]; @@ -595,9 +657,10 @@ template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) int aux0 = aux[0]>aux[1] ? aux[0] : aux[1]; int aux2 = aux[2]>aux[3] ? aux[2] : aux[3]; return aux0>aux2 ? aux0 : aux2; +#endif // EIGEN_VECTORIZE_SSE4_1 } -#if (defined __GNUC__) +#if EIGEN_COMP_GNUC // template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) // { // Packet4f res = b; @@ -705,6 +768,110 @@ struct palign_impl<Offset,Packet2d> }; #endif +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4f,4>& kernel) { + _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2d,2>& kernel) { + __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]); + kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]); + kernel.packet[1] = tmp; +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4i,4>& kernel) { + __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]); + __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]); + __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]); + __m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]); + + kernel.packet[0] = _mm_unpacklo_epi64(T0, T1); + kernel.packet[1] = _mm_unpackhi_epi64(T0, T1); + kernel.packet[2] = _mm_unpacklo_epi64(T2, T3); + kernel.packet[3] = _mm_unpackhi_epi64(T2, T3); +} + +template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) { + const __m128i zero = _mm_setzero_si128(); + const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m128i false_mask = _mm_cmpeq_epi32(select, zero); +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blendv_epi8(thenPacket, elsePacket, false_mask); +#else + return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket)); +#endif +} +template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) { + const __m128 zero = _mm_setzero_ps(); + const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m128 false_mask = _mm_cmpeq_ps(select, zero); +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blendv_ps(thenPacket, elsePacket, false_mask); +#else + return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket)); +#endif +} +template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) { + const __m128d zero = _mm_setzero_pd(); + const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]); + __m128d false_mask = _mm_cmpeq_pd(select, zero); +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blendv_pd(thenPacket, elsePacket, false_mask); +#else + return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket)); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4f pinsertfirst(const Packet4f& a, float b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blend_ps(a,pset1<Packet4f>(b),1); +#else + return _mm_move_ss(a, _mm_load_ss(&b)); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet2d pinsertfirst(const Packet2d& a, double b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blend_pd(a,pset1<Packet2d>(b),1); +#else + return _mm_move_sd(a, _mm_load_sd(&b)); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4f pinsertlast(const Packet4f& a, float b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blend_ps(a,pset1<Packet4f>(b),(1<<3)); +#else + const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x0,0x0,0x0,0xFFFFFFFF)); + return _mm_or_ps(_mm_andnot_ps(mask, a), _mm_and_ps(mask, pset1<Packet4f>(b))); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet2d pinsertlast(const Packet2d& a, double b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blend_pd(a,pset1<Packet2d>(b),(1<<1)); +#else + const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x0,0xFFFFFFFF,0xFFFFFFFF)); + return _mm_or_pd(_mm_andnot_pd(mask, a), _mm_and_pd(mask, pset1<Packet2d>(b))); +#endif +} + +// Scalar path for pmadd with FMA to ensure consistency with vectorized path. +#ifdef __FMA__ +template<> EIGEN_STRONG_INLINE float pmadd(const float& a, const float& b, const float& c) { + return ::fmaf(a,b,c); +} +template<> EIGEN_STRONG_INLINE double pmadd(const double& a, const double& b, const double& c) { + return ::fma(a,b,c); +} +#endif + } // end namespace internal } // end namespace Eigen |