Allison Piper

@cwharris Does the branch you're looking at include https://github.com/NVIDIA/thrust/commit/8f876ba8c66167c44c7e7388a2692f70dcff85fc ? That fixed several similar overflows in tile size calculations.

Thanks for tracking this down, @cwharris! The changes look reasonable to me, though I'm surprised that `static_cast` on the enum value is necessary. Do you have any insight into why...

> It doesn't matter which argument gets cast to uint64_t, just so long as they're both promoted before multiplication. D'oh, I misread the diff originally -- that totally makes sense....

> could we just clean the whole thing up in a go? We could definitely do quite a bit of cleanup in this area. We can consider this for 2.1.

This seems to be related to NVIDIA/nvidia-docker#1026, which mentions that some oddball quoting is needed when `--gpus device=${DEVICES}` specifies multiple devices, and should be something like `--gpus '"device=1,2"'`. I tried...

I tried a few different levels of escapes on those, but they expand weird in the `docker` command: ``` GPU_OPTS="--gpus '\\\"device=${VISIBLE_DEVICES}\\\"'" ``` ``` cuda11.5.1-devel-ubuntu20.04-gcc9: Pulling from gpuci/cccl Digest: sha256:c5c1f5e1534287be3fa48a0896c40b470cfc76abfd6ba6505c437cf131802663 Status:...

Took a quick look at this in the debugger it appears that memory is getting corrupt -- the `parallel_for` algorithm's `tile_base` is consistently set to 0, but is a very...

Ok, just wanted to make sure. I'm planning to spend a bit more time looking at this in case it is in our libraries before we escalate.

I spent a couple more hours looking into this, and things seem to go off the rails around [this line](https://github.com/NVIDIA/thrust/blob/0f2f85cc69d6bc7aacdbe4f73ab9da74b46b5c52/thrust/system/detail/generic/binary_search.inl#L116). The lhs of the assignment (`thrust::get(t)`) is producing a null...

It looks like we'll need to update the docs here, similar to NVIDIA/thrust#1587. Thanks for pointing this out!