Build option DSO=1 generates libverifiable.so which can be
used to reduce the combined binary size.
Build option NAME_SUFFIX can be used to a add suffix to all
generated binaries. e.g. NAME_SUFFIX=_mpi
Added new make target: clean_intermediates
`NCCL_TESTS_SPLIT` serves as new way of computing the color for splitting communicators.
Will be overrided by `NCCL_TESTS_SPLIT_MASK`.
Examples:
NCCL_TESTS_SPLIT_MASK="0x7" # color = rank & 0x7. What we do today to run on a DGX with one GPU per node.
NCCL_TESTS_SPLIT="AND 0x7" # color = rank & 0x7. New way to run on one GPU per node on a DGX, equivalent to NCCL_TESTS_SPLIT_MASK=0x7
NCCL_TESTS_SPLIT="MOD 72" # color = rank % 72. One GPU per NVLink domain on an NVL72 system.
NCCL_TESTS_SPLIT="DIV 72" # color = rank / 72. Intra NVLink domain on NVL72.
You can also use: "%" "&" "|" "/" for short.
Extra spaces in the middle will be automatically ignored.
Not case sensitive.
The followings are all equivalent:
NCCL_TESTS_SPLIT="%0x7"
NCCL_TESTS_SPLIT="%0b111"
NCCL_TESTS_SPLIT="AND 7"
NCCL_TESTS_SPLIT="and 0x7"
Ensure that ncclstringtotype iterates only over data types known to
nccl-tests (as indicated by test_typenum), not over a potentially larger
set of all NCCL types.
as relative to top-level directory. This done is by abspath'ing it before
passing it to subdirectory Makefile's.
The old behavior had two cases: with and without BUILDDIR being set by
the user. With BUILDDIR not set, the build dir would be named "build"
in the top-level directory. If BUILDDIR was set, then the build dir
would be placed at "src/${BUILDDIR}".
The new behavior is simpler, if BUILDDIR is not set then it defaults
to "build", and the directory holding the final build is always at just
"${BUILDDIR}" in the top level.
* 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.
