Fix #1849 by calling device_global_buffer() unconditionally (#1862)

* Hold on to the CUDA buffer - second try
see: https://github.com/davisking/dlib/pull/1855#discussion_r314666348

* Fix #1849 by calling device_global_buffer() unconditionally
This commit is contained in:
Juha Reunanen 2019-08-17 18:01:00 +03:00 committed by Davis E. King
parent b02be8c734
commit d7fb51e60b
1 changed files with 3 additions and 6 deletions

View File

@ -438,16 +438,12 @@ namespace dlib
const size_t bytes_per_plane = subnetwork_output.nr()*subnetwork_output.nc()*sizeof(uint16_t); const size_t bytes_per_plane = subnetwork_output.nr()*subnetwork_output.nc()*sizeof(uint16_t);
// Allocate a cuda buffer to store all the truth images and also one float // Allocate a cuda buffer to store all the truth images and also one float
// for the scalar loss output. // for the scalar loss output.
if (!work)
{
work = device_global_buffer(); work = device_global_buffer();
} buf = work->get(subnetwork_output.num_samples()*bytes_per_plane + sizeof(float));
cuda_data_void_ptr buf = work->get(subnetwork_output.num_samples()*bytes_per_plane + sizeof(float));
cuda_data_void_ptr loss_buf = buf; cuda_data_void_ptr loss_buf = buf;
buf = buf+sizeof(float); buf = buf+sizeof(float);
// copy the truth data into a cuda buffer. // copy the truth data into a cuda buffer.
for (long i = 0; i < subnetwork_output.num_samples(); ++i, ++truth) for (long i = 0; i < subnetwork_output.num_samples(); ++i, ++truth)
{ {
@ -471,6 +467,7 @@ namespace dlib
); );
mutable std::shared_ptr<resizable_cuda_buffer> work; mutable std::shared_ptr<resizable_cuda_buffer> work;
mutable cuda_data_void_ptr buf;
}; };
// ------------------------------------------------------------------------------------ // ------------------------------------------------------------------------------------