jonysy

@paholg have you seen @killercup's latest post? "[A way to customize Rust error messages]" [A way to customize Rust error messages]: https://scribbles.pascalhertleif.de/hook-into-rustc-errors.html

[discussion](http://softwareengineering.stackexchange.com/a/215766/115193): > With OpenCL 2 the compilation model changed some to be much more flexible. Now there is an LLVM backend that targets the new SPIR-V intermediate representation, so any...

In addition: The body of the [`SharedTensor::autosync`] method contains the logic in question. @alexandermorozov's original comment: > Backends may define transfers asymmetrically; for example, CUDA may know how to transfer...

[L48](https://github.com/alexandermorozov/collenchyma-refactor/blob/master/src/lib.rs#L48): > Async operations: it looks like currently most time is spent waiting for in/out transfers even on mid-range GPU hardware. Async may help a lot. Async can be implemented...

@drahnr's point on OpenCL implementation via Gitter: > The way it is currently implemented lacks the either a cl finish or waiting for the last event in the chain (forward...

[A look at GPU memory transfer](http://blog.theincredibleholk.org/blog/2012/11/29/a-look-at-gpu-memory-transfer/)

[Multi-Device Execution][1] > Once a system has multiple devices, there are two main complications: deciding which device to place the com- putation for each node in the graph, and then...

[alexandermorozov](https://github.com/autumnai/collenchyma/issues/58#issuecomment-200849820): > `ndarray` solves first part by accepting constant values, slices, Vecs and iterators as initializers. But it **doesn't allow to allocate uninitialized chunk of memory to be overwritten later**;...

> .. I can easily test things. This project needs all the love it can get! I'll try to set some time aside to work on the CUDA API this...

> I can try and help, but I've only got a GTX460, don't really know if it supports the CUDA version you're interested in supporting, and I don't believe it...