diff options
| author |  Benoit Steiner <benoit.steiner.goog@gmail.com> | 2016-06-09 08:22:27 -0700 |
|---|---|---|
| committer | Benoit Steiner <benoit.steiner.goog@gmail.com> | 2016-06-09 08:22:27 -0700 |
| commit | aa33446dace833fbf06632e586c80119b3d8ac11 (patch) | |
| tree | d7bc59330e69e19de37da9fffc64016ee0febec0 /unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h | |
| parent | 15890c304edbccedc8a989468ed3fc475f428059 (diff) | |
Improved support for vectorization of 16-bit floats
Diffstat (limited to 'unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h')
| -rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h | 8 |
1 files changed, 4 insertions, 4 deletions
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h index e82530955..1b4fdd03f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h @@ -331,7 +331,7 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> { #ifdef EIGEN_HAS_CUDA_FP16 static const bool HasOptimizedImplementation = !Op::IsStateful && (internal::is_same<typename Self::CoeffReturnType, float>::value || - (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && Op::PacketAccess)); + (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess)); #else static const bool HasOptimizedImplementation = !Op::IsStateful && internal::is_same<typename Self::CoeffReturnType, float>::value; @@ -346,7 +346,7 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> { return; } - FullReductionLauncher<Self, Op, OutputType, Op::PacketAccess>::run(self, reducer, device, output, num_coeffs); + FullReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs); } }; @@ -608,7 +608,7 @@ struct InnerReducer<Self, Op, GpuDevice> { #ifdef EIGEN_HAS_CUDA_FP16 static const bool HasOptimizedImplementation = !Op::IsStateful && (internal::is_same<typename Self::CoeffReturnType, float>::value || - (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && Op::PacketAccess)); + (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess)); #else static const bool HasOptimizedImplementation = !Op::IsStateful && internal::is_same<typename Self::CoeffReturnType, float>::value; @@ -627,7 +627,7 @@ struct InnerReducer<Self, Op, GpuDevice> { return true; } - return InnerReductionLauncher<Self, Op, OutputType, Op::PacketAccess>::run(self, reducer, device, output, num_coeffs_to_reduce, num_preserved_vals); + return InnerReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs_to_reduce, num_preserved_vals); } }; |