Andrey Alekseenko
Andrey Alekseenko
> I believe some other exception wording might suite here. That would already be a big improvement. The current error is generic to the point of uselessness. But AMD offers...
> The spec doesn't allow for the sub_group_sizes info descriptor to throw. Are you sure? The spec says, "All member functions of the `device` class are synchronous and errors are...
BTW, if the spec doesn't allow for the sub_group_sizes info descriptor to throw, then we have a similar defect with the host device (but with a more descriptive `sycl::exception::what()`): ```...
Gentle bump. Unless I'm mistaken, this can be a very nasty issue potentially causing sub-group collective operations to produce wrong results.
> It's surprising that LLVM does not manage to realize that it doesn't need the offset (I wonder if this persists with LLVM 13-15). I don't think LLVM can do...
> This does not work because the specification does not allow this: A user kernel can always expect an nd_item as argument, and we cannot pass in some modified type....
On AMD, the impact (in terms of lines of assembly) is even more drastic. ROCm 5.1.3, bundled clang, hip:gfx908 target: ``` s_load_dword s0, s[4:5], 0x4 s_load_dword s11, s[4:5], 0x8 s_load_dwordx2...
> From [install-spirv.md](https://github.com/illuhad/hipSYCL/blob/develop/doc/install-spirv.md), it looks like I must download some Intel software and configure it when compiling hipSYCL. Could I make a dummy backend that just emits LLVM IR or...
> @al42and could you provide the Dockerfile for installing this repo on Linux? By "this repo" you mean hipSYCL with Intel/LLVM? No, we only use them separately. That said, building...
> Trying to download the binary from GitHub releases: I would suggest to use `wget`. `curl` handles redirects poorly. That said, I'm pretty sure the binary is x86_64-only and is...