tree 595484d12e5d53640d23a56bb3dc87c830416bee
parent 027d77c29fd41f2af6dd27f869a0aa2be3268525
author Nicolas Capens <capn@google.com> 1644158959 -0500
committer Nicolas Capens <nicolascapens@google.com> 1644328601 +0000

Unroll transcendentals benchmarks by 10x

This reduces the overhead of the benchmark framework and calling the
routine. For example the results for rr::Cos vs. sw::Cos went from
9.6ns / 75ns to 75ns / 725ns. It illustrates that the timings did not
increase tenfold, and the gap between the two widened so it provides
a better measure of the achieved speedup as we further optimize these
transcendental SPIR-V instructions.

This change also copies the full result to an output buffer, to
ensure LLVM doesn't just compute the first element.

Bug: b/158231104
Change-Id: Ia9cde41b32559b16322d074767e986a68f00706d
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/62348
Kokoro-Result: kokoro <noreply+kokoro@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>