Request for status/documentation update w.r.t. OpenCL

[ThijsWithaar]

When testing on linux, sample code runs fine on host (host_selector), but not on any OpenCL device (cpu_selector). Not sure if it's a mistake on my side, or if it's just not yet supposed to be working yet.

Background info: When I modify any the demo_parallel_matrix_add test to use the 'cpu_selector', I get a terminate called after throwing an instance of 'cl::sycl::non_cl_error'. I have no clue whether that's just not implemented, or if I did something wrong

[keryell]

Actually for now only the host_selector is working, and it is far from reaching official conformance yet. For the OpenCL part, the interoperability mode is working but it is not what you are looking, I guess. :-) There is a prototype with some examples in tests/device_compiler targeting OpenCL with a device compiler mainly for FPGA, with an inefficient compiler scheme for GPU yet. We are in the process to migrate to Clang/LLVM 7 to improve a lot of things but I do not have resources to work on it before November... Sorry for the current state. :-( But at least Codeplay announced recently the first SYCL conformant implementation. While not open source, it works and you can download it for free.

[ThijsWithaar]

For the OpenCL part, the interoperability mode is working ..

Thanks for the clarification. I suspected there were limitations, but wasn't sure whether I was doing anything wrong. It seems not, so I'll just exercise some patience.

But at least Codeplay announced recently the first SYCL conformant implementation

Thanks again. An older 0.4 release didn't work on debian(buster). Just tried their 1.0 release, and that one does work. A bit unrelated, maybe, but triSycl has an .x() accessor in cl::sycl::id<2>, whereas ComputeCpp has a [0] accessor.

[airlied]

I've reworked a bunch of that accessor code in a branch, I should start dequeuing the fixes soon

[airlied]

#189 moves the accessors off the id.