iame6162013

Hmm, the fix you proposed seems fairly complex. If the issue is that coriander doesn't use the same version of llvm as the driver/other(?), why don't we allow for coriander...

Another error ``` + bazel build --verbose_failures @grpc//:grpc_cpp_plugin INFO: Found 1 target... ERROR: /home/iame/.cache/bazel/_bazel_iame/88f8289dc53da1a7a11159bcedf3496d/external/protobuf/BUILD:113:1: undeclared inclusion(s) in rule '@protobuf//:protobuf': this rule is missing dependency declarations for the following files included...

If you want it to use a different llvm you could go into /coriander/install_distro.py and edit target_url in install_llvm, this will cause install_distro.py to download & install a different llvm...

llvm 4.0.1 has some fixes without api changes, coriander doesn't use this atm.

Looking at the code a bit more, if you make coriander use the system installed version, you should also uncomment maybe_install_llvm in main in install_distro.py

> what do you mean by "cuda compliant library"? in what sense do you feel that Coriander is not what you call a "cuda compliant library"? what would you do...

> For these kernels, the question is, how to get them to run on a non-CUDA, non-PTX, non-SASS GPU? No, that would probably be extremely difficult, and as far as...

I think we had a little misunderstanding there, I thought you meant after the default compilation, which coriander doesn't do, I think.

So, if I where to create this host-side API, and compile the kernels with coriander, would I then be capable of ~running~ compiling & running tensorflow without any changes to...

May I ask what the slow part is? Is it saving to disk that is slow, i.e. the disk is >80% utilized, or does the compressing e.g. to png take...