parsec
parsec copied to clipboard
ICLDisco
PaRSEC is a generic framework for architecture aware scheduling and management of micro-tasks on distributed, GPU accelerated, many-core heterogeneous architectures. PaRSEC assigns computation threads...
## Describe the bug Termdet causes recursive to crash. ### To Reproduce Steps to reproduce the behavior: 1. Checkout b958ae9f 2. Checkout parsec 9fc74b6f1 3. Compile with the following options...
The number of active send-recv are hardcoded in defines. This should be read from an mca param. _Originally posted by @abouteiller in https://github.com/ICLDisco/parsec/pull/538#discussion_r1191199059_
Info API
This is the Info key API At this point, I'd like to get feedback on the API (the .h) before I go and implement the ops. In particular, the proposed...
This PCB scheduler uses some bits (MCA defined) of the priority word to define which 'group' of threads can schedule a task with this priority. There is additional documentation in...
## Describe the bug Run `profile2h5` on the most recent version of dplasma and PaRSEC (as of April 25, 2023) and you will notice that it segfaults. With tracing on...
## Describe the bug There is a deadlock when running on GPU. This deadlock happens when using NEW with a large matrix size (memory does not fit into GPU memory)....
@DSMishler and Joseph John have traces with GPU tasks that show a surprising behavior: the GPU tasks are decorated with m, n, k, tcid, etc... but the CPU tasks don't...
## Description When the matrix can not be fitted into GPU memory, performance issues will occur. It's better to find a way to improve the data reuse on GPU, whether...
## Description If the received data type is different from what it needs on GPUs, maybe it's better to do node-level datatype conversion plus caching on devices instead of tile-level.
## Description Right now, the datatype conversion needs to be dealt with explicitly. It would be great to find a way to convert this data type in runtime. The direct...