devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

issue #2581: review and cleanup of compiler version checks

Browse Source
This commit is contained in:
Sean McBride
2023-01-17 18:58:34 +00:00
committed by Rasmus Munk Larsen
parent b523120687
commit d70b4864d9
25 changed files with 103 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@@ -670,7 +670,7 @@ template<> EIGEN_STRONG_INLINE Packet8i pcmp_eq(const Packet8i& a, const Packet8
}
template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const Packet8f& b) {
#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
#if EIGEN_GNUC_STRICT_LESS_THAN(6,3,0)
// There appears to be a bug in GCC, by which the optimizer may flip
// the argument order in calls to _mm_min_ps/_mm_max_ps, so we have to
// resort to inline ASM here. This is supposed to be fixed in gcc6.3,
@@ -684,7 +684,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const
#endif
}
template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const Packet4d& b) {
#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
#if EIGEN_GNUC_STRICT_LESS_THAN(6,3,0)
// See pmin above
Packet4d res;
asm("vminpd %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b));
@@ -705,7 +705,7 @@ template<> EIGEN_STRONG_INLINE Packet8i pmin<Packet8i>(const Packet8i& a, const
}
template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) {
#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
#if EIGEN_GNUC_STRICT_LESS_THAN(6,3,0)
// See pmin above
Packet8f res;
asm("vmaxps %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b));
@@ -716,7 +716,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const
#endif
}
template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) {
#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
#if EIGEN_GNUC_STRICT_LESS_THAN(6,3,0)
// See pmin above
Packet4d res;
asm("vmaxpd %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b));

4
Eigen/src/Core/arch/AVX512/PacketMath.h
View File

@@ -31,7 +31,7 @@ namespace internal {
#endif
// Disable the code for older versions of gcc that don't support many of the required avx512 math instrinsics.
#if EIGEN_GNUC_AT_LEAST(5, 3) || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC >= 1923 || EIGEN_COMP_ICC >= 1900
#if EIGEN_GNUC_STRICT_AT_LEAST(5,3,0) || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC >= 1923 || EIGEN_COMP_ICC >= 1900
#define EIGEN_HAS_AVX512_MATH 1
#else
#define EIGEN_HAS_AVX512_MATH 0
@@ -2384,7 +2384,7 @@ EIGEN_STRONG_INLINE Packet16f Bf16ToF32(const Packet16bf& a) {
EIGEN_STRONG_INLINE Packet16bf F32ToBf16(const Packet16f& a) {
Packet16bf r;
#if defined(EIGEN_VECTORIZE_AVX512BF16) && EIGEN_GNUC_AT_LEAST(10, 1)
#if defined(EIGEN_VECTORIZE_AVX512BF16) && EIGEN_GNUC_STRICT_AT_LEAST(10,1,0)
// Since GCC 10.1 supports avx512bf16 and C style explicit cast
// (C++ static_cast is not supported yet), do conversion via intrinsic
// and register path for performance.

2
Eigen/src/Core/arch/AVX512/PacketMathFP16.h
View File

@@ -17,7 +17,7 @@ namespace Eigen {
namespace internal {
// Disable the code for older versions of gcc that don't support many of the required avx512 math instrinsics.
#if EIGEN_GNUC_AT_LEAST(5, 3) || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC >= 1923 || EIGEN_COMP_ICC >= 1900
#if EIGEN_GNUC_STRICT_AT_LEAST(5,3,0) || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC >= 1923 || EIGEN_COMP_ICC >= 1900
#define EIGEN_HAS_AVX512_MATH 1
#else
#define EIGEN_HAS_AVX512_MATH 0

4
Eigen/src/Core/arch/AltiVec/MathFunctions.h
View File

@@ -61,7 +61,7 @@ Packet4f patan<Packet4f>(const Packet4f& _x)
}
#ifdef EIGEN_VECTORIZE_VSX
#ifndef EIGEN_COMP_CLANG
#if !EIGEN_COMP_CLANG
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet4f prsqrt<Packet4f>(const Packet4f& x)
{
@@ -103,7 +103,7 @@ template<> EIGEN_STRONG_INLINE Packet8bf psqrt<Packet8bf> (const Packet8bf& a){
BF16_TO_F32_UNARY_OP_WRAPPER(psqrt<Packet4f>, a);
}
#ifndef EIGEN_COMP_CLANG
#if !EIGEN_COMP_CLANG
template<> EIGEN_STRONG_INLINE Packet8bf prsqrt<Packet8bf> (const Packet8bf& a){
BF16_TO_F32_UNARY_OP_WRAPPER(prsqrt<Packet4f>, a);
}

3
Eigen/src/Core/arch/AltiVec/PacketMath.h
View File

@@ -2935,8 +2935,7 @@ template<> EIGEN_STRONG_INLINE Packet2d psignbit(const Packet2d& a) { return (P
// are buggy, https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70963
template<>
inline Packet2l pcast<Packet2d, Packet2l>(const Packet2d& x) {
#if EIGEN_GNUC_AT_LEAST(5, 4) || \
(EIGEN_GNUC_AT(6, 1) && __GNUC_PATCHLEVEL__ >= 1)
#if EIGEN_GNUC_STRICT_AT_LEAST(7,1,0)
return vec_cts(x, 0); // TODO: check clang version.
#else
double tmp[2];

2
Eigen/src/Core/arch/GPU/Complex.h
View File

@@ -39,7 +39,7 @@
// To allow std::complex operator use on device, define _OVERRIDE_COMPLEX_SPECIALIZATION_
// prior to first inclusion of <complex>. This prevents ICC from adding
// its own specializations, so our custom ones below can be used instead.
#if !(defined(EIGEN_COMP_ICC) && defined(_USE_COMPLEX_SPECIALIZATION_))
#if !(EIGEN_COMP_ICC && defined(_USE_COMPLEX_SPECIALIZATION_))
// Import Eigen's internal operator specializations.
#define EIGEN_USING_STD_COMPLEX_OPERATORS \

4
Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
View File

@@ -96,7 +96,7 @@ struct gebp_traits <float,float,false,false,Architecture::NEON,GEBPPacketFull>
template<int LaneID>
EIGEN_STRONG_INLINE void madd_helper(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c) const
{
#if EIGEN_COMP_GNUC_STRICT
#if EIGEN_GNUC_STRICT_LESS_THAN(9,0,0)
// 1. workaround gcc issue https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89101
// vfmaq_laneq_f32 is implemented through a costly dup, which was fixed in gcc9
// 2. workaround the gcc register split problem on arm64-neon
@@ -166,7 +166,7 @@ struct gebp_traits <double,double,false,false,Architecture::NEON>
template <int LaneID>
EIGEN_STRONG_INLINE void madd_helper(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c) const
{
#if EIGEN_COMP_GNUC_STRICT
#if EIGEN_GNUC_STRICT_LESS_THAN(9,0,0)
// 1. workaround gcc issue https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89101
// vfmaq_laneq_f64 is implemented through a costly dup, which was fixed in gcc9
// 2. workaround the gcc register split problem on arm64-neon

6
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@@ -3696,11 +3696,11 @@ template<> EIGEN_STRONG_INLINE Packet4bf pnegate<Packet4bf>(const Packet4bf& a)
// Clang 3.5 in the iOS toolchain has an ICE triggered by NEON intrisics for double.
// Confirmed at least with __apple_build_version__ = 6000054.
#ifdef __apple_build_version__
#if EIGEN_COMP_CLANGAPPLE
// Let's hope that by the time __apple_build_version__ hits the 601* range, the bug will be fixed.
// https://gist.github.com/yamaya/2924292 suggests that the 3 first digits are only updated with
// major toolchain updates.
#define EIGEN_APPLE_DOUBLE_NEON_BUG (__apple_build_version__ < 6010000)
#define EIGEN_APPLE_DOUBLE_NEON_BUG (EIGEN_COMP_CLANGAPPLE < 6010000)
#else
#define EIGEN_APPLE_DOUBLE_NEON_BUG 0
#endif
@@ -3928,7 +3928,7 @@ template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
// Other reduction functions:
// mul
#if EIGEN_COMP_CLANG && defined(__apple_build_version__)
#if EIGEN_COMP_CLANGAPPLE
template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
{ return (vget_low_f64(a) * vget_high_f64(a))[0]; }
#else

8
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@@ -475,7 +475,7 @@ template<> EIGEN_STRONG_INLINE Packet16b pcmp_eq(const Packet16b& a, const Packe
template<> EIGEN_STRONG_INLINE Packet4i pcmp_le(const Packet4i& a, const Packet4i& b) { return por(pcmp_lt(a,b), pcmp_eq(a,b)); }
template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) {
#if EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC < 63
#if EIGEN_GNUC_STRICT_LESS_THAN(6,3,0)
// There appears to be a bug in GCC, by which the optimizer may
// flip the argument order in calls to _mm_min_ps, so we have to
// resort to inline ASM here. This is supposed to be fixed in gcc6.3,
@@ -494,7 +494,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const
#endif
}
template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) {
#if EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC < 63
#if EIGEN_GNUC_STRICT_LESS_THAN(6,3,0)
// There appears to be a bug in GCC, by which the optimizer may
// flip the argument order in calls to _mm_min_pd, so we have to
// resort to inline ASM here. This is supposed to be fixed in gcc6.3,
@@ -525,7 +525,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const
template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) {
#if EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC < 63
#if EIGEN_GNUC_STRICT_LESS_THAN(6,3,0)
// There appears to be a bug in GCC, by which the optimizer may
// flip the argument order in calls to _mm_max_ps, so we have to
// resort to inline ASM here. This is supposed to be fixed in gcc6.3,
@@ -544,7 +544,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const
#endif
}
template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) {
#if EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC < 63
#if EIGEN_GNUC_STRICT_LESS_THAN(6,3,0)
// There appears to be a bug in GCC, by which the optimizer may
// flip the argument order in calls to _mm_max_pd, so we have to
// resort to inline ASM here. This is supposed to be fixed in gcc6.3,