simd-alphatensor-rs VS stdarch

thanks for the feedback u/lebensterben. Please see https://github.com/drbh/simd-alphatensor-rs#benchmarks for the results and https://github.com/drbh/simd-alphatensor-rs/blob/main/benches/my_benchmark/main.rs for the actual tests. Overall the results are promising - however, benches should always be taken with a grain of salt. I hope this is helpful!

Good to know! How would you compare it with the std_detect implementation, underlying the standard library (https://github.com/rust-lang/stdarch/tree/master/crates/std_detect)?

Fortunately, @burntsushi (thanks) is sharing the workaround in the rust-lang github issue. The trick is to reinterpret a (non-tagged) union for this.

The stdarch tests will test target-specific stuff like SIMD while to UI tests are more generic. They both test that the compiler will produce an executable that produce the correct output (or that rustc itself will produce the correct output).

I made enough progress to be able to compile the tests of stdarch and most of them pass: that's a huge milestone!

You can check the code. Apparently the std implementation uses the OSXSAVE register to check that the OS supports saving AVX/AVX2 registers during context switches and only then enables it. In a non-std context one might not generally be able to depend on the OSXSAVE register.