mirror of
https://github.com/NVIDIA/nccl-tests.git
synced 2026-05-01 11:58:19 +08:00
51af5572bf
* 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.
12 lines
575 B
Makefile
12 lines
575 B
Makefile
# We requires both of the following paths to be set upon including this makefile
|
|
# TEST_VERIFIABLE_SRCDIR = <points to this directory>
|
|
# TEST_VERIFIABLE_BUILDDIR = <points to destination of .o file>
|
|
|
|
TEST_VERIFIABLE_HDRS = $(TEST_VERIFIABLE_SRCDIR)/verifiable.h
|
|
TEST_VERIFIABLE_OBJS = $(TEST_VERIFIABLE_BUILDDIR)/verifiable.o
|
|
|
|
$(TEST_VERIFIABLE_BUILDDIR)/verifiable.o: $(TEST_VERIFIABLE_SRCDIR)/verifiable.cu $(TEST_VERIFY_REDUCE_HDRS)
|
|
@printf "Compiling %s\n" $@
|
|
@mkdir -p $(TEST_VERIFIABLE_BUILDDIR)
|
|
$(NVCC) -o $@ $(NVCUFLAGS) -c $(TEST_VERIFIABLE_SRCDIR)/verifiable.cu
|