Michal Babej

OK, was mistaken a bit, it happened on the old 2g microSD also, only much later. Possibly related: i remember i changed some variables in /proc/sys/vm (`dirty_background_bytes`, `dirty_bytes`, `dirty_expire_centisecs`, `dirty_writeback_centisecs`)...

> One test fails here with iGPU (both LZ and CL) and Intel CPU: It works for me on my iGPU (with both LZ and CL), so i'm going to...

PoCL CPU device has FP16 support only when compiled with LLVM 16 and higher (and that support is quite incomplete).

IIRC there were multiple reasons. The major one was that the Mali does not support FP64 and there is no emulation (AFAIK). Some of the tests have "int" or "unsigned"...

Removing the barriers (+using event dependencies) significantly reduced the difference (to ~4x slower), but there was also a kernel problem - SYCL was using fast-math by default, and the kernels...

@pjaaskel no, not yet.

> The purpose of these tests was to make sure that we have all the implementations for the device library.... you wouldn't know which function you were missing. makes sense...

@pjaaskel FYI tested SPIR-V from chipStar 17 directly (`llvm-spirv -r xyz.spv`); with LLVM 17 SPIRV-Translator it doesn't crash, LLVM 16 SPIRV-Translator crashes on debuginfo with the same backtrace.

@Kerilk so should i cherry-pick this into my PR, or do you plan to create a PR from the `dispatch` branch ?

`LD_LIBRARY_PATH` is only used when loading a library (by `ld`); it's not used when GCC or other compiler searches for a library at build time. CMake has its [own variable](https://cmake.org/cmake/help/latest/variable/CMAKE_LIBRARY_PATH.html)...