diff options
author | Artem Belevich <tra@google.com> | 2017-11-07 14:00:29 -0800 |
---|---|---|
committer | TensorFlower Gardener <gardener@tensorflow.org> | 2017-11-07 14:06:32 -0800 |
commit | c070ab2a58ab5dc66da54db2daac4fa8e9f191e9 (patch) | |
tree | b8c8c4eeb0e076f46ba8b6c0a855883d920aa07b /tensorflow/stream_executor/stream_executor_pimpl.h | |
parent | 7183348f3270b7f9c1b333970e4f9abf6b3c4d8a (diff) |
Clean up kernels cached by CUDAExecutor.
Notify CUDA executor on deallocation of previously loaded GPUExecutable and
unload corresponding module when all executables that were using it are gone.
PiperOrigin-RevId: 174908193
Diffstat (limited to 'tensorflow/stream_executor/stream_executor_pimpl.h')
-rw-r--r-- | tensorflow/stream_executor/stream_executor_pimpl.h | 3 |
1 files changed, 3 insertions, 0 deletions
diff --git a/tensorflow/stream_executor/stream_executor_pimpl.h b/tensorflow/stream_executor/stream_executor_pimpl.h index 9c225e5fae..66c50d47e9 100644 --- a/tensorflow/stream_executor/stream_executor_pimpl.h +++ b/tensorflow/stream_executor/stream_executor_pimpl.h @@ -104,6 +104,9 @@ class StreamExecutor { // platform, false is returned. bool GetKernel(const MultiKernelLoaderSpec &spec, KernelBase *kernel); + // Releases any state associated with the previously loaded kernel. + void UnloadKernel(const KernelBase *kernel); + // Synchronously allocates an array on the device of type T with element_count // elements. template <typename T> |