GitLab

Our stalls, hangs, lockups, whatever you want to call them are caused by the
mcnt processing logic and how blocks are initialized and pushed thorugh the
pipeline. This pipeline was based off the paper correlator and when FLAG had
started development the paper net thread was used as a starting point.

This reverts back to some of the old paper code ideas with some modifications
(like the block info object being a variable passed around at main instead of
inn the process packet function) and some of the re-initialization.

In its current state the logic seems to work great for the beamformer but the
correlator is now broken after two scans. This is most likely due to how blocks
are being filled, set free but I am not sure if it is on the net thread side or
the correlator side.

This commit is marking this as WIP and then going to move onto a different paper
repo that has a revised mcnt processing logic that we want to try. That revised
logic is a newer version of the paper code and interesting enough in their git
commit logs they mention that they are seeing hangs as well.

We may not ever use these but were extended from the base hashpipe structure
when experimenting with improving the net thread processing logic.

Also added checking for run_threads() in the wait free loop in corsave thread.
This lets the thread detect if a signal was received and can exit. Might be a
better way to do this but works for now.

Decreasse indent and clean up code in correlator thread for readability

Without the condition to check if the thread should be running (e.g., the call
to run_threads()) the save thread hangs waiting for blocks to be filled and does
not register the signal interrupt correctly.

Minor file formatting and clean up to help readability while going through to
read and debug the code

Finish making the install script accept a different install location. There is a
weird bug where when the beamformer is not installed to the production site of
/home/groups/flag/hash that the Makefile incorrectly uses the wrong CUDA path.
This was overcome temporarily by passing the $CUDA path into the Makefile.

Pinned memory, streams and events implemented.

See merge request !1

Most recent version of code used on telescope with setVersion script and soft links for lib and src to the versions folder

This contains versions of the scalloping and no scalloping versions of the pfb. The version installed in this commit is the scalloping one. To install versions, use rsync -a and copy the appropriate directories (lib and src)  over.

Just updating everything. Change made to Makefile for cuda JIT issue. It's the first commented line of code.

Made changes to transpose and databuf.h. In transpose thread, I changed line 185 to ACQUIRE, and in flag_databuf.h, I changed VERBOSE to 1 and DEBUG to 1.

New New code

forgot to track the new beamform thread

Merging modified matlab files

Mid July/Aug Commissioning Improvements

This commit was made at the end of me leaving GBO during the middle of the July/Aug commissioning. Many of these updates have improved the performance of the pipeline. Performance measured in the frequency of stalled/hanging banks and bad blocks. We tuned core assignments, input block counts and more. Decreasesd the amount of time accessed by shared memory. Prevented packets sneaking through at the end of scans in the net thread. Removed the cpu transpose for the beamformer and implemented on the GPU.

Fixes to pfb corr mode. The GPU needed to flush the buffer that stores the last filter length worth of data. Also added new clean up logic lat waits until the net_thread has gone to the idle state. Also added a method for most threads that looks at the amount of locked blocks extending a hashpipe function

Updates from May commissioning

These commits contain the changes I made in the last few days at the May commissioning. I had gone through and made changes to hashpipe and the clean up structure for the threads. Before, hashpipe would temrinate the child process immediately, now the children terminate themselves and clean up then report to parent process. Also, hashpipe now clears shared memory and reclaims everything at the end and checks at the beginning. This commit also contains the fix to the PFB mode. The traspose thread to reduce data trasfered was indexing out of bounds.

Added while loops around semaphore-checking calls, which, for some reason, timeout instead of blocking until the semaphore is released...