GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-11 19:29:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

Enable the conversion between floats and half floats on older GPUs that support it.

Browse Source
This commit is contained in:
Benoit Steiner 2016-03-03 10:33:20 -08:00
parent 2de8cc9122
commit 1da10a7358
1 changed files with 5 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -21,7 +21,7 @@ struct scalar_cast_op<float, half> {
EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
typedef half result_type; typedef half result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half operator() (const float& a) const { EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half operator() (const float& a) const {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
return __float2half(a); return __float2half(a);
#else #else
assert(false && "tbd"); assert(false && "tbd");
@ -40,7 +40,7 @@ struct scalar_cast_op<int, half> {
EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
typedef half result_type; typedef half result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half operator() (const int& a) const { EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half operator() (const int& a) const {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
return __float2half(static_cast<float>(a)); return __float2half(static_cast<float>(a));
#else #else
assert(false && "tbd"); assert(false && "tbd");
@ -59,7 +59,7 @@ struct scalar_cast_op<half, float> {
EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
typedef float result_type; typedef float result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const half& a) const { EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const half& a) const {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
return __half2float(a); return __half2float(a);
#else #else
assert(false && "tbd"); assert(false && "tbd");
@ -85,7 +85,7 @@ struct type_casting_traits<half, float> {
}; };
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<half2, float4>(const half2& a, const half2& b) { template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<half2, float4>(const half2& a, const half2& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
float2 r1 = __half22float2(a); float2 r1 = __half22float2(a);
float2 r2 = __half22float2(b); float2 r2 = __half22float2(b);
return make_float4(r1.x, r1.y, r2.x, r2.y); return make_float4(r1.x, r1.y, r2.x, r2.y);
@ -106,7 +106,7 @@ struct type_casting_traits<float, half> {
template<> EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) { template<> EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) {
// Simply discard the second half of the input // Simply discard the second half of the input
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
return __float22half2_rn(make_float2(a.x, a.y)); return __float22half2_rn(make_float2(a.x, a.y));
#else #else
assert(false && "tbd"); assert(false && "tbd");