smaslov-intel

> This commit adds support for handling invalid kernels. > > Sometimes kernel cannot be generated for specific platform, when trying to append launch of such faulty kernel exception is...

> Kernel's compiled from spirv using fp64 on platforms without fp64 support are poisoned (code is replaced) with printf. This is done to enable building module with fp64 kernels on...

> That's how intel-graphics-compiler works. The problem is if user tries to run this kernel. It would be much better to get an error instead of executing said kernel and...

> > Wouldn't it be much more friendly to give an error at AOT compilation-time (and request JIT compilation) rather than deferring to run-time? > > This is a must...

> We can use single spirv with multiple kernels (including those using fp64) and compile it to native binary for every device. But you said just earlier that you can't...

> If user adds attribute "reqd_work_group_size(X,Y,Z)" then it's expected for work group size to be (X, Y, Z) (any other work group size on setGroupSize call will result in an...

> > If user adds attribute "reqd_work_group_size(X,Y,Z)" then it's expected for work group size to be (X, Y, Z) (any other work group size on setGroupSize call will result in...

Why is this not relying on the existing device-cache? I would expect that the new "device_impl" would not be created because of this code: https://github.com/intel/llvm/blob/f6420c769d473c00be736b8df361e10dd98a6009/sycl/source/detail/platform_impl.cpp#L219 > If a device is...

> that `createSyclObjFromImpl` does not call the `Platform->getOrMakeDeviceImpl` method that checks the device-cache. This is why we needed to call `device::get_devices` that does check the device-cache and as far as...

> `MDeviceCache` is a private member of `platform_impl` so I thought it could be controversial to let `make_device` access it directly? It is true that if `make_device` were to access...