BUG with nvshmem 3.2.5 for bitcode compiling

Accelerated Computing GPU-Accelerated Libraries
1

for nvshem 3.2.5, when compiling bitcode we found a BUG in nvshmem/src/include/non_abi/device/common/nvshmemi_common_device.cuh, which makes libnvshmem_device.bc can’t handle memcpy with src % 16 == 0 and dst % 16 == 0 and size % 16 != 0.

the BUG is shown as below:

image
image1920×730 142 KB

2 Likes
2

You are correct, thank you. I have added a fix to the next release and called out the issue in our changelog.

1 Like