diff options
Diffstat (limited to 'Eigen/src/Core/arch/NEON/PacketMath.h')
-rw-r--r-- | Eigen/src/Core/arch/NEON/PacketMath.h | 233 |
1 files changed, 224 insertions, 9 deletions
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 0504c095c..9afd86bec 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -20,14 +20,24 @@ namespace internal { #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 #endif +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#endif + +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#endif + // FIXME NEON has 16 quad registers, but since the current register allocator // is so bad, it is much better to reduce it to 8 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS -#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8 +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 #endif +typedef float32x2_t Packet2f; typedef float32x4_t Packet4f; typedef int32x4_t Packet4i; +typedef int32x2_t Packet2i; typedef uint32x4_t Packet4ui; #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ @@ -39,7 +49,7 @@ typedef uint32x4_t Packet4ui; #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ const Packet4i p4i_##NAME = pset1<Packet4i>(X) -#if defined(__llvm__) && !defined(__clang__) +#if EIGEN_COMP_LLVM && !EIGEN_COMP_CLANG //Special treatment for Apple's llvm-gcc, its NEON packet types are unions #define EIGEN_INIT_NEON_PACKET2(X, Y) {{X, Y}} #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {{X, Y, Z, W}} @@ -52,11 +62,11 @@ typedef uint32x4_t Packet4ui; // arm64 does have the pld instruction. If available, let's trust the __builtin_prefetch built-in function // which available on LLVM and GCC (at least) -#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || defined(__GNUC__) +#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC #define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR); #elif defined __pld #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR) -#elif !defined(__aarch64__) +#elif !EIGEN_ARCH_ARM64 #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ( " pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" ); #else // by default no explicit prefetching @@ -66,11 +76,12 @@ typedef uint32x4_t Packet4ui; template<> struct packet_traits<float> : default_packet_traits { typedef Packet4f type; - typedef Packet4f half; + typedef Packet2f half; enum { Vectorizable = 1, AlignedOnScalar = 1, size = 4, + HasHalfPacket=1, HasDiv = 1, // FIXME check the Has* @@ -84,16 +95,17 @@ template<> struct packet_traits<float> : default_packet_traits template<> struct packet_traits<int> : default_packet_traits { typedef Packet4i type; - typedef Packet4i half; + typedef Packet2i half; enum { Vectorizable = 1, AlignedOnScalar = 1, - size=4 + size=4, + HasHalfPacket=1 // FIXME check the Has* }; }; -#if EIGEN_GNUC_AT_MOST(4,4) && !defined(__llvm__) +#if EIGEN_GNUC_AT_MOST(4,4) && !EIGEN_COMP_LLVM // workaround gcc 4.2, 4.3 and 4.4 compilatin issue EIGEN_STRONG_INLINE float32x4_t vld1q_f32(const float* x) { return ::vld1q_f32((const float32_t*)x); } EIGEN_STRONG_INLINE float32x2_t vld1_f32 (const float* x) { return ::vld1_f32 ((const float32_t*)x); } @@ -136,6 +148,9 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { +#if EIGEN_ARCH_ARM64 + return vdivq_f32(a,b); +#else Packet4f inv, restep, div; // NEON does not offer a divide instruction, we have to do a reciprocal approximation @@ -154,14 +169,27 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const div = vmulq_f32(a, inv); return div; +#endif } + template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) { eigen_assert(false && "packet integer division are not supported by NEON"); return pset1<Packet4i>(0); } -// for some weird raisons, it has to be overloaded for packet of integers +#ifdef __ARM_FEATURE_FMA +// See bug 936. +// FMA is available on VFPv4 i.e. when compiling with -mfpu=neon-vfpv4. +// FMA is a true fused multiply-add i.e. only 1 rounding at the end, no intermediate rounding. +// MLA is not fused i.e. does 2 roundings. +// In addition to giving better accuracy, FMA also gives better performance here on a Krait (Nexus 4): +// MLA: 10 GFlop/s ; FMA: 12 GFlops/s. +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vfmaq_f32(c,a,b); } +#else template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vmlaq_f32(c,a,b); } +#endif + +// No FMA instruction for int, so use MLA unconditionally. template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return vmlaq_s32(c,a,b); } template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vminq_f32(a,b); } @@ -472,6 +500,193 @@ ptranspose(PacketBlock<Packet4i,4>& kernel) { kernel.packet[3] = vcombine_s32(vget_high_s32(tmp1.val[1]), vget_high_s32(tmp2.val[1])); } +//---------- double ---------- +#if EIGEN_ARCH_ARM64 + +#if (EIGEN_COMP_GNUC_STRICT && defined(__ANDROID__)) || defined(__apple_build_version__) +// Bug 907: workaround missing declarations of the following two functions in the ADK +__extension__ static __inline uint64x2_t __attribute__ ((__always_inline__)) +vreinterpretq_u64_f64 (float64x2_t __a) +{ + return (uint64x2_t) __a; +} + +__extension__ static __inline float64x2_t __attribute__ ((__always_inline__)) +vreinterpretq_f64_u64 (uint64x2_t __a) +{ + return (float64x2_t) __a; +} +#endif + +typedef float64x2_t Packet2d; +typedef float64x1_t Packet1d; + +template<> struct packet_traits<double> : default_packet_traits +{ + typedef Packet2d type; + typedef Packet1d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + HasHalfPacket=1, + + HasDiv = 1, + // FIXME check the Has* + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 0, + HasSqrt = 0 + }; +}; + +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2}; typedef Packet2d half; }; + +template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return vdupq_n_f64(from); } + +template<> EIGEN_STRONG_INLINE Packet2d plset<double>(const double& a) +{ + Packet2d countdown = EIGEN_INIT_NEON_PACKET2(0, 1); + return vaddq_f64(pset1<Packet2d>(a), countdown); +} +template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return vaddq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return vsubq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return vnegq_f64(a); } + +template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vmulq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vdivq_f64(a,b); } + +#ifdef __ARM_FEATURE_FMA +// See bug 936. See above comment about FMA for float. +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vfmaq_f64(c,a,b); } +#else +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vmlaq_f64(c,a,b); } +#endif + +template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vminq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vmaxq_f64(a,b); } + +// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics +template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f64(from); } + +template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f64(from); } + +template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) +{ + return vld1q_dup_f64(from); +} +template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f64(to, from); } + +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f64(to, from); } + +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, DenseIndex stride) +{ + Packet2d res; + res = vsetq_lane_f64(from[0*stride], res, 0); + res = vsetq_lane_f64(from[1*stride], res, 1); + return res; +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, DenseIndex stride) +{ + to[stride*0] = vgetq_lane_f64(from, 0); + to[stride*1] = vgetq_lane_f64(from, 1); +} +template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ARM_PREFETCH(addr); } + +// FIXME only store the 2 first elements ? +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(a, 0); } + +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return vcombine_f64(vget_high_f64(a), vget_low_f64(a)); } + +template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vabsq_f64(a); } + +#if EIGEN_COMP_CLANG && defined(__apple_build_version__) +// workaround ICE, see bug 907 +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return (vget_low_f64(a) + vget_high_f64(a))[0]; } +#else +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return vget_lane_f64(vget_low_f64(a) + vget_high_f64(a), 0); } +#endif + +template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) +{ + float64x2_t trn1, trn2; + + // NEON zip performs interleaving of the supplied vectors. + // We perform two interleaves in a row to acquire the transposed vector + trn1 = vzip1q_f64(vecs[0], vecs[1]); + trn2 = vzip2q_f64(vecs[0], vecs[1]); + + // Do the addition of the resulting vectors + return vaddq_f64(trn1, trn2); +} +// Other reduction functions: +// mul +#if EIGEN_COMP_CLANG && defined(__apple_build_version__) +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return (vget_low_f64(a) * vget_high_f64(a))[0]; } +#else +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return vget_lane_f64(vget_low_f64(a) * vget_high_f64(a), 0); } +#endif + +// min +template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(vpminq_f64(a, a), 0); } + +// max +template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(vpmaxq_f64(a, a), 0); } + +// this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors, +// see bug 347 and this LLVM bug: http://llvm.org/bugs/show_bug.cgi?id=11074 +#define PALIGN_NEON(Offset,Type,Command) \ +template<>\ +struct palign_impl<Offset,Type>\ +{\ + EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\ + {\ + if (Offset!=0)\ + first = Command(first, second, Offset);\ + }\ +};\ + +PALIGN_NEON(0,Packet2d,vextq_f64) +PALIGN_NEON(1,Packet2d,vextq_f64) +#undef PALIGN_NEON + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2d,2>& kernel) { + float64x2_t trn1 = vzip1q_f64(kernel.packet[0], kernel.packet[1]); + float64x2_t trn2 = vzip2q_f64(kernel.packet[0], kernel.packet[1]); + + kernel.packet[0] = trn1; + kernel.packet[1] = trn2; +} +#endif // EIGEN_ARCH_ARM64 + } // end namespace internal } // end namespace Eigen |