[CI] Enable FPGA on postcommit for Linux

[uditagarwal97]

This PR enables testing on opencl::fpga for Linux in postcommit.
I have currently disabled the tests that were failing on the fpga and created another Github issue for the same: #12683. I am not very sure if the failures are because of some implementation bug or if that feature is not supported on fpga.

[Alcpz]

I think I found out the issue. We instantiate kernels with both 32bit and 64 bit datatypes. When the JIT compilation gets called the first time, all kernels get incorporated to the same module for compilation, triggering the issue.

-fsycl-device-code-split=per_kernel splits each kernel into independent modules, which should fix the issue.

Edit: @aelovikov-intel the query is working as intended.

[aelovikov-intel]

I think I found out the issue. We instantiate kernels with both 32bit and 64 bit datatypes. When the JIT compilation gets called the first time, all kernels get incorporated to the same module for compilation, triggering the issue.

-fsycl-device-code-split=per_kernel splits each kernel into independent modules, which should fix the issue.

Edit: @aelovikov-intel the query is working as intended.

That should be handled by Optional Kernel Features (https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OptionalDeviceFeatures.md), shouldn't it? In particular, I'd expect two device images created even without that option. I admit that I haven't worked closely with that feature, so maybe I'm missing something.

+ @AlexeySachkov @steffenlarsen

[steffenlarsen]

@aelovikov-intel is right, the module-splitting done by sycl-post-commit should split kernels using atomic64 into a separate module. Since the failure seems to be coming from the online compiler and not from our runtime, which it should, I suspect we have a case of atomics on 64-bit values that aren't appropriately marked as [[__sycl_detail__::__uses_aspects__(aspect::atomic64)]].

I suggest you open an issue here on Github with it, add per_kernel splitting to the test, and add a comment that it's a bug due to the newly opened issue.

[bader]

accelerator or fpga?

[uditagarwal97]

I think it's accelerator only because that's what the name of the feature is (https://github.com/intel/llvm/blob/sycl/sycl/test-e2e/lit.cfg.py#L678) when fpga device is reported by sycl-ls.

[uditagarwal97]

The precommit failure in ESIMD/unified_memory_api/scatter_lacc_dg2_pvc.cpp seems unrealted: #12769