|
Uni20
Asynchronous tensor and runtime library.
|
Configuration for one scoped process-wide CUDA runtime installation. More...
#include <runtime.hpp>
Public Attributes | |
| std::optional< int > | default_device {} |
| Device used by lookup without an explicit ordinal; defaults to the first enrolled device. | |
| std::vector< int > | device_ordinals {} |
| CUDA device ordinals to enroll; an empty list enrolls every visible device. | |
| unsigned int | stream_flags = cudaStreamNonBlocking |
| CUDA stream creation flags used by every enrolled device. | |
| std::size_t | streams_per_device = 8 |
| Number of actually-idle streams owned by each enrolled device. | |
Configuration for one scoped process-wide CUDA runtime installation.
| std::optional<int> uni20::cuda::RuntimeConfig::default_device {} |
Device used by lookup without an explicit ordinal; defaults to the first enrolled device.
| std::vector<int> uni20::cuda::RuntimeConfig::device_ordinals {} |
CUDA device ordinals to enroll; an empty list enrolls every visible device.
| unsigned int uni20::cuda::RuntimeConfig::stream_flags = cudaStreamNonBlocking |
CUDA stream creation flags used by every enrolled device.
| std::size_t uni20::cuda::RuntimeConfig::streams_per_device = 8 |
Number of actually-idle streams owned by each enrolled device.