UseFloat16¶
Enables the use of 16 bit floating point computations (half precision). This can speed up some computations significantly, if your GPU supports half arithmetic. At the moment, the only command that can use half computations is ComputeDisparityMap with the matching method set to Correlation .
Note
This needs a GPU with a compute capability of at least 5.3. If your GPU does not support half computations, we will silently fall back to using normal floats.
Note
The performance of half computations highly depends on your specific device. Some GPU models have very limited support even if they have a high enough compute capability (e.g. the GeForce 10 series). On these devices, enabling this option will make the matching significantly slower than the normal float version. To find out whether your device supports FP16, you can check the list of Nvidia GPUs on Wikipedia. To benefit from half computations, the value in the column “Half precision” should be higher than the value in the column “Single precision”.
Format
Boolean |
Whether the commands are allowed to use half computations. |
Default
false ( true on Nvidia Jetson)