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

Merged eigen/eigen into default

Browse Source
This commit is contained in:
Konstantinos Margaritis
2017-08-24 12:24:01 +03:00
parent 4ce5ec5197 21633e585b
commit 1affe3d8df
83 changed files with 620 additions and 288 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -16,7 +16,7 @@ namespace Eigen {
namespace internal {
#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
// Many std::complex methods such as operator+, operator-, operator* and
// operator/ are not constexpr. Due to this, clang does not treat them as device

30
Eigen/src/Core/arch/CUDA/Half.h
View File

@@ -140,7 +140,7 @@ struct half : public half_impl::half_base {
namespace half_impl {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
// Intrinsics for native fp16 support. Note that on current hardware,
// these are no faster than fp32 arithmetic (you need to use the half2
@@ -281,7 +281,7 @@ union FP32 {
};
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
return __float2half(ff);
#elif defined(EIGEN_HAS_FP16_C)
@@ -336,7 +336,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h) {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
return __half2float(h);
#elif defined(EIGEN_HAS_FP16_C)
@@ -370,7 +370,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) {
return (a.x & 0x7fff) == 0x7c00;
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
return __hisnan(a);
#else
return (a.x & 0x7fff) > 0x7c00;
@@ -386,7 +386,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
return result;
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
return half(hexp(a));
#else
return half(::expf(float(a)));
@@ -396,7 +396,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half expm1(const half& a) {
return half(numext::expm1(float(a)));
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
#if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
#if defined(EIGEN_HAS_CUDA_FP16) && EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
return half(::hlog(a));
#else
return half(::logf(float(a)));
@@ -409,7 +409,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
return half(::log10f(float(a)));
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
return half(hsqrt(a));
#else
return half(::sqrtf(float(a)));
@@ -431,14 +431,14 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
return half(::tanhf(float(a)));
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
return half(hfloor(a));
#else
return half(::floorf(float(a)));
#endif
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300
return half(hceil(a));
#else
return half(::ceilf(float(a)));
@@ -446,7 +446,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
return __hlt(b, a) ? b : a;
#else
const float f1 = static_cast<float>(a);
@@ -455,7 +455,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
#endif
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
return __hlt(a, b) ? b : a;
#else
const float f1 = static_cast<float>(a);
@@ -576,7 +576,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
return Eigen::half(::expf(float(a)));
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
#if EIGEN_CUDACC_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530
return Eigen::half(::hlog(a));
#else
return Eigen::half(::logf(float(a)));
@@ -610,14 +610,14 @@ struct hash<Eigen::half> {
// Add the missing shfl_xor intrinsic
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
}
#endif
// ldg() has an overload for __half, but we also need one for Eigen::half.
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
return Eigen::half_impl::raw_uint16_to_half(
__ldg(reinterpret_cast<const unsigned short*>(ptr)));
@@ -625,7 +625,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr)
#endif
#if defined(__CUDA_ARCH__)
#if defined(EIGEN_CUDA_ARCH)
namespace Eigen {
namespace numext {

2
Eigen/src/Core/arch/CUDA/MathFunctions.h
View File

@@ -17,7 +17,7 @@ namespace internal {
// Make sure this is only available when targeting a GPU: we don't want to
// introduce conflicts between these packet_traits definitions and the ones
// we'll use on the host side (SSE, AVX, ...)
#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
float4 plog<float4>(const float4& a)
{

10
Eigen/src/Core/arch/CUDA/PacketMath.h
View File

@@ -17,7 +17,7 @@ namespace internal {
// Make sure this is only available when targeting a GPU: we don't want to
// introduce conflicts between these packet_traits definitions and the ones
// we'll use on the host side (SSE, AVX, ...)
#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
#if defined(EIGEN_CUDACC) && defined(EIGEN_USE_GPU)
template<> struct is_arithmetic<float4> { enum { value = true }; };
template<> struct is_arithmetic<double2> { enum { value = true }; };
@@ -196,7 +196,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<double>(double* to
template<>
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Aligned>(const float* from) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
return __ldg((const float4*)from);
#else
return make_float4(from[0], from[1], from[2], from[3]);
@@ -204,7 +204,7 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Aligned>(const fl
}
template<>
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Aligned>(const double* from) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
return __ldg((const double2*)from);
#else
return make_double2(from[0], from[1]);
@@ -213,7 +213,7 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Aligned>(const
template<>
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Unaligned>(const float* from) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
return make_float4(__ldg(from+0), __ldg(from+1), __ldg(from+2), __ldg(from+3));
#else
return make_float4(from[0], from[1], from[2], from[3]);
@@ -221,7 +221,7 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Unaligned>(const
}
template<>
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Unaligned>(const double* from) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
return make_double2(__ldg(from+0), __ldg(from+1));
#else
return make_double2(from[0], from[1]);

28
Eigen/src/Core/arch/CUDA/PacketMathHalf.h
View File

@@ -15,7 +15,7 @@ namespace Eigen {
namespace internal {
// Most of the following operations require arch >= 3.0
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDACC__) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
template<> struct is_arithmetic<half2> { enum { value = true }; };
@@ -69,7 +69,7 @@ template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half*
template<>
__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
#if __CUDA_ARCH__ >= 350
#if EIGEN_CUDA_ARCH >= 350
return __ldg((const half2*)from);
#else
return __halves2half2(*(from+0), *(from+1));
@@ -78,7 +78,7 @@ template<>
template<>
__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
#if __CUDA_ARCH__ >= 350
#if EIGEN_CUDA_ARCH >= 350
return __halves2half2(__ldg(from+0), __ldg(from+1));
#else
return __halves2half2(*(from+0), *(from+1));
@@ -116,7 +116,7 @@ ptranspose(PacketBlock<half2,2>& kernel) {
}
template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
return __halves2half2(a, __hadd(a, __float2half(1.0f)));
#else
float f = __half2float(a) + 1.0f;
@@ -125,7 +125,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half&
}
template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
return __hadd2(a, b);
#else
float a1 = __low2float(a);
@@ -139,7 +139,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, cons
}
template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
return __hsub2(a, b);
#else
float a1 = __low2float(a);
@@ -153,7 +153,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, cons
}
template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
return __hneg2(a);
#else
float a1 = __low2float(a);
@@ -165,7 +165,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
return __hmul2(a, b);
#else
float a1 = __low2float(a);
@@ -179,7 +179,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, cons
}
template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
return __hfma2(a, b, c);
#else
float a1 = __low2float(a);
@@ -225,7 +225,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, cons
}
template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
return __hadd(__low2half(a), __high2half(a));
#else
float a1 = __low2float(a);
@@ -235,7 +235,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2&
}
template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
__half first = __low2half(a);
__half second = __high2half(a);
return __hgt(first, second) ? first : second;
@@ -247,7 +247,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const ha
}
template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
__half first = __low2half(a);
__half second = __high2half(a);
return __hlt(first, second) ? first : second;
@@ -259,7 +259,7 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const ha
}
template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
#if __CUDA_ARCH__ >= 530
#if EIGEN_CUDA_ARCH >= 530
return __hmul(__low2half(a), __high2half(a));
#else
float a1 = __low2float(a);
@@ -284,7 +284,7 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
return __floats2half2_rn(r1, r2);
}
#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
template<> __device__ EIGEN_STRONG_INLINE
half2 plog<half2>(const half2& a) {

8
Eigen/src/Core/arch/CUDA/TypeCasting.h
View File

@@ -19,7 +19,7 @@ struct scalar_cast_op<float, Eigen::half> {
EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
typedef Eigen::half result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
return __float2half(a);
#else
return Eigen::half(a);
@@ -37,7 +37,7 @@ struct scalar_cast_op<int, Eigen::half> {
EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
typedef Eigen::half result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
return __float2half(static_cast<float>(a));
#else
return Eigen::half(static_cast<float>(a));
@@ -55,7 +55,7 @@ struct scalar_cast_op<Eigen::half, float> {
EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
typedef float result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
return __half2float(a);
#else
return static_cast<float>(a);
@@ -69,7 +69,7 @@ struct functor_traits<scalar_cast_op<Eigen::half, float> >
#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
template <>
struct type_casting_traits<Eigen::half, float> {