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

updates based on PR feedback

Browse Source 
There are two major changes (and a few minor ones which are not listed here...see PR discussion for details)

1. Eigen::half implementations for HIP and CUDA have been merged.
This means that
- `CUDA/Half.h` and `HIP/hcc/Half.h` got merged to a new file `GPU/Half.h`
- `CUDA/PacketMathHalf.h` and `HIP/hcc/PacketMathHalf.h` got merged to a new file `GPU/PacketMathHalf.h`
- `CUDA/TypeCasting.h` and `HIP/hcc/TypeCasting.h` got merged to a new file `GPU/TypeCasting.h`

After this change the `HIP/hcc` directory only contains one file `math_constants.h`. That will go away too once that file becomes a part of the HIP install.

2. new macros EIGEN_GPUCC, EIGEN_GPU_COMPILE_PHASE and EIGEN_HAS_GPU_FP16 have been added and the code has been updated to use them where appropriate.
- `EIGEN_GPUCC` is the same as `(EIGEN_CUDACC || EIGEN_HIPCC)`
- `EIGEN_GPU_DEVICE_COMPILE` is the same as `(EIGEN_CUDA_ARCH || EIGEN_HIP_DEVICE_COMPILE)`
- `EIGEN_HAS_GPU_FP16` is the same as `(EIGEN_HAS_CUDA_FP16 or EIGEN_HAS_HIP_FP16)`
This commit is contained in:
Deven Desai
2018-06-14 10:21:54 -04:00
parent ba972fb6b4
commit b6cc0961b1
29 changed files with 425 additions and 185 deletions
Download Patch File Download Diff File Expand all files Collapse all files

145
Eigen/src/Core/arch/GPU/PacketMathHalf.h
View File

@@ -7,15 +7,16 @@
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_PACKET_MATH_HALF_CUDA_H
#define EIGEN_PACKET_MATH_HALF_CUDA_H
#ifndef EIGEN_PACKET_MATH_HALF_GPU_H
#define EIGEN_PACKET_MATH_HALF_GPU_H
namespace Eigen {
namespace internal {
// Most of the following operations require arch >= 3.0
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
(defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIPCC) && defined(EIGEN_HIP_DEVICE_COMPILE))
template<> struct is_arithmetic<half2> { enum { value = true }; };
@@ -43,7 +44,18 @@ template<> struct packet_traits<Eigen::half> : default_packet_traits
template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; };
template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
#if defined(EIGEN_HAS_OLD_HIP_FP16)
return half2half2(from);
#else
return __half2half2(from);
#endif
#else // EIGEN_CUDA_ARCH
return __half2half2(from);
#endif
}
template<> __device__ EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
@@ -69,20 +81,46 @@ 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 defined(EIGEN_HIP_DEVICE_COMPILE)
#if defined(EIGEN_HAS_OLD_HIP_FP16)
return __halves2half2((*(from+0)), (*(from+1)));
#else
return __ldg((const half2*)from);
#endif
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 350
return __ldg((const half2*)from);
#else
return __halves2half2(*(from+0), *(from+1));
#endif
#endif
}
template<>
__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
#if defined(EIGEN_HAS_OLD_HIP_FP16)
return __halves2half2((*(from+0)), (*(from+1)));
#else
return __halves2half2(__ldg(from+0), __ldg(from+1));
#endif
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 350
return __halves2half2(__ldg(from+0), __ldg(from+1));
#else
return __halves2half2(*(from+0), *(from+1));
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
@@ -117,15 +155,29 @@ ptranspose(PacketBlock<half2,2>& kernel) {
}
template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
return __halves2half2(a, __hadd(a, __float2half(1.0f)));
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
return __halves2half2(a, __hadd(a, __float2half(1.0f)));
#else
float f = __half2float(a) + 1.0f;
return __halves2half2(a, __float2half(f));
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
return __hadd2(a, b);
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
return __hadd2(a, b);
#else
@@ -137,9 +189,17 @@ template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, cons
float r2 = a2 + b2;
return __floats2half2_rn(r1, r2);
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
return __hsub2(a, b);
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
return __hsub2(a, b);
#else
@@ -151,9 +211,17 @@ template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, cons
float r2 = a2 - b2;
return __floats2half2_rn(r1, r2);
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
return __hneg2(a);
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
return __hneg2(a);
#else
@@ -161,11 +229,19 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
float a2 = __high2float(a);
return __floats2half2_rn(-a1, -a2);
#endif
#endif
}
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 defined(EIGEN_HIP_DEVICE_COMPILE)
return __hmul2(a, b);
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
return __hmul2(a, b);
#else
@@ -177,9 +253,17 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, cons
float r2 = a2 * b2;
return __floats2half2_rn(r1, r2);
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
return __hfma2(a, b, c);
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
return __hfma2(a, b, c);
#else
@@ -193,9 +277,21 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, con
float r2 = a2 * b2 + c2;
return __floats2half2_rn(r1, r2);
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
#if defined(EIGEN_HAS_OLD_HIP_FP16)
return h2div(a, b);
#else
return __h2div(a, b);
#endif
#else // EIGEN_CUDA_ARCH
float a1 = __low2float(a);
float a2 = __high2float(a);
float b1 = __low2float(b);
@@ -203,6 +299,8 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, cons
float r1 = a1 / b1;
float r2 = a2 / b2;
return __floats2half2_rn(r1, r2);
#endif
}
template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
@@ -226,6 +324,12 @@ 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 defined(EIGEN_HIP_DEVICE_COMPILE)
return __hadd(__low2half(a), __high2half(a));
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
return __hadd(__low2half(a), __high2half(a));
#else
@@ -233,9 +337,19 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2&
float a2 = __high2float(a);
return Eigen::half(__float2half(a1 + a2));
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
__half first = __low2half(a);
__half second = __high2half(a);
return __hgt(first, second) ? first : second;
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
__half first = __low2half(a);
__half second = __high2half(a);
@@ -245,9 +359,19 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const ha
float a2 = __high2float(a);
return a1 > a2 ? __low2half(a) : __high2half(a);
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
__half first = __low2half(a);
__half second = __high2half(a);
return __hlt(first, second) ? first : second;
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
__half first = __low2half(a);
__half second = __high2half(a);
@@ -257,9 +381,17 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const ha
float a2 = __high2float(a);
return a1 < a2 ? __low2half(a) : __high2half(a);
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
#if defined(EIGEN_HIP_DEVICE_COMPILE)
return __hmul(__low2half(a), __high2half(a));
#else // EIGEN_CUDA_ARCH
#if EIGEN_CUDA_ARCH >= 530
return __hmul(__low2half(a), __high2half(a));
#else
@@ -267,6 +399,8 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const ha
float a2 = __high2float(a);
return Eigen::half(__float2half(a1 * a2));
#endif
#endif
}
template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
@@ -285,7 +419,8 @@ template<> __device__ EIGEN_STRONG_INLINE half2 pexpm1<half2>(const half2& a) {
return __floats2half2_rn(r1, r2);
}
#if EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530
#if (EIGEN_CUDACC_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530) || \
defined(EIGEN_HIP_DEVICE_COMPILE)
template<> __device__ EIGEN_STRONG_INLINE
half2 plog<half2>(const half2& a) {
@@ -1130,4 +1265,4 @@ ptranspose(PacketBlock<Packet4h,4>& kernel) {
}
}
#endif // EIGEN_PACKET_MATH_HALF_CUDA_H
#endif // EIGEN_PACKET_MATH_HALF_GPU_H