GitLab

profiling: improving NVTX profiling CPU+GPU execution

See merge request !5

nvtx+summary: improving nvtx by tracing initialization, which also fixes potential bogus first measurement of Projection for the summary

bugfix (from valgrind): even when CC is not used, it should be initialized to a…

See merge request !4

To enable it use undisclosed CMake option -DUSE_NVTX=ON (by default it is OFF)

bugfix (from valgrind): even when CC is not used, it should be initialized to a value (e.g. 0?) as it is printed later. Also some new code developments might use this CC value (by mistake), so better to keep it initialized

doc and copyright update

See merge request !3

minor fixes

See merge request !2

REVERT: temporarily disabled fastest CUDA detection, since it strangely causes error on dvl machine (reverted from commit 1459580e)

adding older preliminary data for the initial Autotuning algorithm. Results obtained on dvl01 machine

WATCH OUT! This is additional CUDA timing profiling activated by setting BIOEM_DEBUG_OUTPUT=4. However, this profiling is quite intrusive, as it adds additional synchronizations between GPUs and OMP that are now sequentially working on maps comparison. If the BIOEM_DEBUG_OUTPUT<4, code is ignored and the performance is back to normal