| 1 |
# Cuda library |
| 2 |
|
| 3 |
This library should be linked to by getting cflags and libs from |
| 4 |
gstreamer-cuda-{{ gst_api_version.md }}.pc |
| 5 |
|
| 6 |
> NOTE: This library API is considered *unstable* |
| 7 |
|
| 8 |
## Environment variables |
| 9 |
|
| 10 |
The GStreamer CUDA library inspects following environment variables |
| 11 |
|
| 12 |
**`GST_CUDA_CRITICAL_ERRORS`. (Since: 1.24)** |
| 13 |
|
| 14 |
This environment variable can be set to a comma-separated list of CUresult |
| 15 |
values (see CUDA driver API documentation). GStreamer CUDA library will |
| 16 |
abort when the user registered error is detected. This environment can be useful |
| 17 |
when unrecoverable CUDA error happens. Thus in-process error recovery |
| 18 |
(e.g., relaunching new pipeline) is not expected to work, and therefore |
| 19 |
the process should be relaunched. |
| 20 |
|
| 21 |
Example: `GST_CUDA_CRITICAL_ERRORS=2,700` |
| 22 |
|
| 23 |
As a result of the above example, if `CUDA_ERROR_OUT_OF_MEMORY(2)` or |
| 24 |
`CUDA_ERROR_ILLEGAL_ADDRESS(700)` error is detected in GStreamer CUDA library, |
| 25 |
the process will be aborted. |
| 26 |
|
| 27 |
|
| 28 |
**`GST_CUDA_ENABLE_STREAM_ORDERED_ALLOC`. (Since: 1.26)** |
| 29 |
|
| 30 |
As of 1.26, GStreamer CUDA library supports stream ordered CUDA allocation |
| 31 |
(e.g., cuMemAllocAsync). The new allocation method is disabled by default |
| 32 |
unless it's explicitly requested via buffer pool option. |
| 33 |
This environment variable can be used to change the default behavior |
| 34 |
so that the stream ordered allocation can be used by default. |