computationalradiationphysics / cracen Goto Github PK

See https://github.com/ComputationalRadiationPhysics/genericCommunicator/blob/dev/CMakeLists.txt as example

• add errorhandler to cuda calls

currently in file Dsp.cu

• When input file is empty, don't use a hardcoded default
• instead, give an error or (even better) show the help!

Chunks of input data could be distributed with some kind of master/slave principle #MPI

Add a python api.

all tasks of a block write in the global memory at FitFunction.hpp: getWindow() to initialize the window to zero.

Add a precomputation and postcomputation Hook to cracen.

Boost 1.61. has a nice implementation.
http://www.boost.org/doc/libs/1_61_0/doc/html/circular_buffer/rationale.html

Add a way to add meta data to the payload data. The way the meta data is sent should be configurable to take other paths, to increase bandwidth and decrease latancy of the payload data.

All install dependencies and run dependiencies should be listed
All important versions should be documented

• Change SendPolcy Api to not need cage
  -> this is better anyway, since there is no valid cage ref in the beginning and the lambda call in constructor of cracen is a mess
• make template arguments to arguments of the constructor
• benefit from lambdas on creation

Ideas:

• use different format
  • should be well documented (VTK? 👅, binary?)
  • should be space efficient
• stay with JSON, but compress it before writing to disk
  • either compress directly in C++ with the use of a library
  • or supply a command-line option to write the output to stdout. Then, you can use a pipe to compress the stream on the fly with arbitrary software (tar)

More parallel (if(threadIdx.x == 0 ) {...} )

e.g. here

Some dependencies are missing in the Makefile.
Workaround: run make clean before make.

fix all failing tests

Float overflow during calculation. Calculation in LeveMarq.hpp AT*A (X^7)^2 FOR X=0..SAMPLE_COUNT too big for floats.
MAX order = log (SAMPLE_COUNT, sqrt(MAX_FLOAT))

Redefine x domain to [0..1]. But questionable resolution. => use mixed precision

Compiling on measurement pc uses to many registers for LevenbergMarquard kernel (>100) on other machines the kernel only needs 40 registers

The grpc update of graybat lead to different config objects needed for zmq contexts. Update Benchmark, Tests, and all the examples to be compatible with the signaling server.

If a cracen gets deleted, it tries to clean all of the allocated resources. However some threads are waiting on synchronus or asynchronus send or receive operations from graybat. Since there is no way to stop these threads either via graybat or via the std::thread interface, these may be stuck forever.

This should not be a problem, since the destructor is not called in continguous data processing (cracen instances are living over the whole programs lifespan) and even if some zombie threads are alive, they don't take much memory and nexto zero cpu time.

One problem might arise, if one zombie decides to wake up and acts on the released data. This would most likely lead to seg fault or undefined behaviour.

Add an option to stream incomeing and outgoing messages to hard drives as a backup.

Implement a c api.

The datarate of the scope reader is to low to utilize even a single GPU. The dumpfiles also proviedes pretty small datasets.
To get useful benchmarks for the fittings a new "compute" node could be added, that multiplys every data set e.g. hundred times before sending it to the fitting node. That way the datastream will supply enough bandwidth to utilize one or more devices or nodes.

Especially to get help output on how to use the applicaiton.

http://www.boost.org/doc/libs/1_56_0/doc/html/program_options.html

Also used in PIConGPU

GaussJordan allocates new shared memory on every iteration.

GaussJordan.hpp: gaussJordan() gets called multiple times (implicit solver) and should not allocated more and more shared mem.