* Added "verifiable", a suite of kernels for generating and verifying reduction
input and output arrays in a bit-precise way.
* Data corruption errors now reported in number of wrong elements instead of max
deviation.
* Use ncclGetLastError.
* Don't run hypercube on non-powers of 2 ranks.
* Fix to hypercube data verification.
* Use "thread local" as the defaut CUDA capture mode.
* Replaced pthread_yield -> sched_yield()
* Bugfix to the cpu-side barrier/allreduce implementations.
Add error checking for minbytes and maxbytes arguments
Also accept lowercase literals when parsing size arguments and print errors and usage on stderr.
Fix#45 : All_gather and reduce_scatter algorithm bandwidth was
computed as time/count*(nranks-1) which is not consistent with the
way we compute it for other collectives.
This change makes algbw higher; busbw is unchanged.
In some cases, the MPI library is not in $(MPI_HOME)/lib but
in $(MPI_HOME)/lib64. For example, on RedHat like Linux system
(CentOS, Amazon Linux), and MPI is installed by yum or rpm.
Under such circumstance, the current make file will cause failure.
This patch address this issue by adding -L$(MPI_HOME)/lib64 to
NVLDFLAGS in src/Makefile.
Signed-off-by: Wei Zhang <wzam@amazon.com>
Major rework to merge most of the changes from the NCCL internal
tests into the public ones
Added "-m <agg_iters>" operation aggregation option.
Data integrity checking is now much more performant at scale.
Startup times at scale are improved.
Test latency units are now displayed in usec.
alan.souza
David Addison
Jithin Jose
Sylvain Jeaugey
John Bachan
Greg Inozemtsev
Luke Yeager
Luke Yeager
Wei Zhang