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

Fix GPU build failures.

Browse Source
This commit is contained in:
Antonio Sanchez
2025-03-09 17:04:41 -07:00
parent 6a4a0b66bd
commit 952eda443b
7 changed files with 37 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
Eigen/src/Core/arch/CUDA/Half.h
View File

@@ -210,13 +210,13 @@ namespace half_impl {
// conversion steps back and forth.
EIGEN_STRONG_INLINE __device__ half operator + (const half& a, const half& b) {
return __hadd(a, b);
return __hadd(static_cast<__half>(a), static_cast<__half>(b));
}
EIGEN_STRONG_INLINE __device__ half operator * (const half& a, const half& b) {
return __hmul(a, b);
return __hmul(static_cast<__half>(a), static_cast<__half>(b));
}
EIGEN_STRONG_INLINE __device__ half operator - (const half& a, const half& b) {
return __hsub(a, b);
return __hsub(static_cast<__half>(a), static_cast<__half>(b));
}
EIGEN_STRONG_INLINE __device__ half operator / (const half& a, const half& b) {
float num = __half2float(a);
@@ -224,7 +224,7 @@ EIGEN_STRONG_INLINE __device__ half operator / (const half& a, const half& b) {
return __float2half(num / denom);
}
EIGEN_STRONG_INLINE __device__ half operator - (const half& a) {
return __hneg(a);
return __hneg(static_cast<__half>(a));
}
EIGEN_STRONG_INLINE __device__ half& operator += (half& a, const half& b) {
a = a + b;
@@ -243,22 +243,22 @@ EIGEN_STRONG_INLINE __device__ half& operator /= (half& a, const half& b) {
return a;
}
EIGEN_STRONG_INLINE __device__ bool operator == (const half& a, const half& b) {
return __heq(a, b);
return __heq(static_cast<__half>(a), static_cast<__half>(b));
}
EIGEN_STRONG_INLINE __device__ bool operator != (const half& a, const half& b) {
return __hne(a, b);
return __hne(static_cast<__half>(a), static_cast<__half>(b));
}
EIGEN_STRONG_INLINE __device__ bool operator < (const half& a, const half& b) {
return __hlt(a, b);
return __hlt(static_cast<__half>(a), static_cast<__half>(b));
}
EIGEN_STRONG_INLINE __device__ bool operator <= (const half& a, const half& b) {
return __hle(a, b);
return __hle(static_cast<__half>(a), static_cast<__half>(b));
}
EIGEN_STRONG_INLINE __device__ bool operator > (const half& a, const half& b) {
return __hgt(a, b);
return __hgt(static_cast<__half>(a), static_cast<__half>(b));
}
EIGEN_STRONG_INLINE __device__ bool operator >= (const half& a, const half& b) {
return __hge(a, b);
return __hge(static_cast<__half>(a), static_cast<__half>(b));
}
#else // Emulate support for half floats
@@ -667,15 +667,15 @@ __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor_sync(unsigned mask, Eigen:
#else // CUDA SDK < 9.0
__device__ EIGEN_STRONG_INLINE Eigen::half __shfl(Eigen::half var, int srcLane, int width = warpSize) {
return static_cast<Eigen::half>(__shfl(static_cast<float>(var), laneMask, width));
return static_cast<Eigen::half>(__shfl(static_cast<float>(var), srcLane, width));
}
__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_up(Eigen::half var, unsigned int delta, int width = warpSize) {
return static_cast<Eigen::half>(__shfl_up(static_cast<float>(var), laneMask, width));
return static_cast<Eigen::half>(__shfl_up(static_cast<float>(var), delta, width));
}
__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_down(Eigen::half var, unsigned int delta, int width = warpSize) {
return static_cast<Eigen::half>(__shfl_down(static_cast<float>(var), laneMask, width));
return static_cast<Eigen::half>(__shfl_down(static_cast<float>(var), delta, width));
}
__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width = warpSize) {

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

@@ -197,7 +197,7 @@ 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
return __ldg((const float4*)from);
return __ldg(reinterpret_cast<const float4*>(from));
#else
return make_float4(from[0], from[1], from[2], from[3]);
#endif
@@ -205,7 +205,7 @@ 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
return __ldg((const double2*)from);
return __ldg(reinterpret_cast<const double2*>(from));
#else
return make_double2(from[0], from[1]);
#endif