Bart Massey

The code in filter-orig.rs here came from this Reddit thread. It apparently is an implementation of a piece of a DSP filter bank.

The Rust version here is my rewrite of that code to clean it up a bit. The C version was hand-translated from the rewritten Rust version. Both versions include very minimal tests that nothing was harmed in the process.

Build with make. Please check the Makefile for CPU type settings and C compiler options, and check the Cargo.toml for other Rust stuff. make clean to clean up, make test to run the tests: an assertion will happen if a test fails.

To summarize my results:

• My Rust rewrite produced better code than the original.

• rustc and clang-12 produced an essentially identical inner loop.

• GCC 10.2 managed to use VFMA and VNFMA instructions to improve on the LLVM result somewhat, although the C standard maybe possibly does not allow this by default due IEEE standards… stuff. Use the fma feature with the Rust code to get a version that uses f32::mul_add() to get essentially the GCC inner loop.

• None of the compilers managed to vectorize the code, which in retrospect is not surprising given its data dependencies in the inner loop.

• On my recent machine (znver3) the inner loop is not unrolled. On my older machine (haswell) the inner loop is unrolled. The difference might be that the compiler CPU model detects no performance advantage from unrolling with the newer CPU.