Attachment 514 Details for Bug 907 – Two patches--new gcc version macro and NEON workaround

This bugzilla service is closed. All entries have been migrated to https://gitlab.com/libeigen/eigen

[patch] Two patches--new gcc version macro and NEON workaround

eigen_macro.patch (text/plain), 4.75 KB, created by Tim Murray on 2015-01-06 00:57:32 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Tim Murray

Created: 2015-01-06 00:57:32 UTC

Size: 4.75 KB

patch

obsolete

># HG changeset patch
># User Tim Murray <timmurray@google.com>
># Date 1420501426 28800
>#      Mon Jan 05 15:43:46 2015 -0800
># Node ID 5c57aecfa8d281e624c7e50f9c65e2232ff5f666
># Parent  80ff8707665ab092dda76db312c18310b8473c98
>Add macro to determine gcc version at patch level granularity.
>
>diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
>--- a/Eigen/src/Core/util/Macros.h
>+++ b/Eigen/src/Core/util/Macros.h
>@@ -102,25 +102,31 @@
> 
> /// \internal EIGEN_GNUC_STRICT set to 1 if the compiler is really GCC and not a compatible compiler (e.g., ICC, clang, mingw, etc.)
> #if EIGEN_COMP_GNUC && !(EIGEN_COMP_CLANG || EIGEN_COMP_CLANG || EIGEN_COMP_MINGW || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM )
>   #define EIGEN_COMP_GNUC_STRICT 1
> #else
>   #define EIGEN_COMP_GNUC_STRICT 0
> #endif
> 
>-
> #if EIGEN_COMP_GNUC
>   #define EIGEN_GNUC_AT_LEAST(x,y) ((__GNUC__==x && __GNUC_MINOR__>=y) || __GNUC__>x)
>   #define EIGEN_GNUC_AT_MOST(x,y)  ((__GNUC__==x && __GNUC_MINOR__<=y) || __GNUC__<x)
>   #define EIGEN_GNUC_AT(x,y)       ( __GNUC__==x && __GNUC_MINOR__==y )
>+  #define EIGEN_GNUC_AT_LEAST(x,y,z) ((__GNUC__==x && __GNUC_MINOR__>=y && __GNUC_PATCHLEVEL__>=z ) || __GNUC__>x)
>+  #define EIGEN_GNUC_AT_MOST(x,y,z)  ((__GNUC__==x && __GNUC_MINOR__<=y && __GNUC_PATCHLEVEL__<=z ) || __GNUC__<x)
>+  #define EIGEN_GNUC_AT(x,y,z)       ( __GNUC__==x && __GNUC_MINOR__==y && __GNUC_PATCHLEVEL__==z )
>+
> #else
>   #define EIGEN_GNUC_AT_LEAST(x,y) 0
>   #define EIGEN_GNUC_AT_MOST(x,y)  0
>   #define EIGEN_GNUC_AT(x,y)       0
>+  #define EIGEN_GNUC_AT_LEAST(x,y,z) 0
>+  #define EIGEN_GNUC_AT_MOST(x,y,z)  0
>+  #define EIGEN_GNUC_AT(x,y,z)       0
> #endif
> 
> // FIXME: could probably be removed as we do not support gcc 3.x anymore
> #if EIGEN_COMP_GNUC && (__GNUC__ <= 3)
> #define EIGEN_GCC3_OR_OLDER 1
> #else
> #define EIGEN_GCC3_OR_OLDER 0
> #endif
># HG changeset patch
># User Tim Murray <timmurray@google.com>
># Date 1420501778 28800
>#      Mon Jan 05 15:49:38 2015 -0800
># Node ID 246d74b505d4f8d3e51af99aa229aff88e81f5d9
># Parent  5c57aecfa8d281e624c7e50f9c65e2232ff5f666
>Add vector float64x1-friendly implementations for AArch64.
>
>gcc 4.9.0 through 4.9.2 incorrectly typedef'd float64x1_t to a
>scalar float64_t, while gcc 4.9.3+ and clang typedef this to a
>1-vector float64_t as specified by the AAPCS64 ABI. This adds
>a workaround to allow gcc 4.9.0 through 4.9.2 to work while
>fixing clang and future gcc versions as well.
>
>diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
>--- a/Eigen/src/Core/arch/NEON/PacketMath.h
>+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
>@@ -592,34 +592,45 @@ template<> EIGEN_STRONG_INLINE void pref
> 
> // 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); }
> 
>+// FIXME: gcc 4.9.2 and below handle float64x1_t as a float64_t, not a vector of float64_t
>+// clang always handles it as a vector
>+#if EIGEN_GNUC_PATCH_AT_MOST(4,9,2)
> template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return vget_low_f64(a) + vget_high_f64(a); }
>+#else
>+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return (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
>+// FIXME: gcc 4.9.2 and below handle float64x1_t as a float64_t, not a vector of float64_t
>+// clang always handles it as a vector
>+#if EIGEN_GNUC_PATCH_AT_MOST(4,9,2)
> template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return vget_low_f64(a) * vget_high_f64(a); }
>-
>+#else
>+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return (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

Actions: View | Diff

Attachments on bug 907: 514