Make "benchmarks game" results return

[kud1ing]

http://benchmarksgame.alioth.debian.org/u64q/rust.php does not show any results anymore. AFAIK this was initally due to a licensing situation, which was resolved in #14248

What needs to be done in order to make the results come back?

[TeXitoi]

I've proposed our relicensed benchmarks:

• https://alioth.debian.org/tracker/index.php?func=detail&aid=314749&group_id=100815&atid=413122
• https://alioth.debian.org/tracker/index.php?func=detail&aid=314750&group_id=100815&atid=413122
• https://alioth.debian.org/tracker/index.php?func=detail&aid=314751&group_id=100815&atid=413122
• https://alioth.debian.org/tracker/index.php?func=detail&aid=314752&group_id=100815&atid=413122
• https://alioth.debian.org/tracker/index.php?func=detail&aid=314753&group_id=100815&atid=413122
• https://alioth.debian.org/tracker/index.php?func=detail&aid=314754&group_id=100815&atid=413122

[kud1ing]

5 benchmarks seem to be back:

• http://benchmarksgame.alioth.debian.org/u32/rust.php
• http://benchmarksgame.alioth.debian.org/u64q/rust.php
• http://benchmarksgame.alioth.debian.org/u32q/rust.php
• http://benchmarksgame.alioth.debian.org/u64/rust.php

[kud1ing]

@TeXitoi: Appearantly there is an unexpected ouput for http://benchmarksgame.alioth.debian.org/u64q/benchmark.php?test=mandelbrot&lang=rust&id=1&data=u64q#log

[TeXitoi]

cc @Ryman

[TeXitoi]

@kud1ing meteor contest, that does not appear in the table, is also back. nbody is not our version.

[Ryman]

@TeXitoi Mandelbrot should be ok again now on master

[TeXitoi]

Most of the benchmarks are back for 0.12. fasta-redux just don't work so I didn't submitted it. mandelbrot gives bad output on x86.

@Ryman do you have an idea about this problem: http://benchmarksgame.alioth.debian.org/u32/program.php?test=mandelbrot&lang=rust&id=1#log

SIMD bug?

[Ryman]

@TeXitoi I looked into this, and cross compiled it a number of times on my machine. Seems like sse2 isn't being used for the x86 build and the floating point calculations are using x87 (which leads to 80bit precision for intermediates rather than 64bit as expected by the benchmark).

I've gotten it to work locally by adding -C target-feature=+sse2 to the rustc arguments, you should be able to add that to the benchmark to get the correct output.

[TeXitoi]

@Ryman thanks for the investigation! I've asked to add these command line flags https://alioth.debian.org/tracker/index.php?func=detail&aid=314839&group_id=100815&atid=413122

[TeXitoi]

@Ryman and congrats to your mandelbrot implementation: 2nd on 4 core amd64 http://benchmarksgame.alioth.debian.org/u64q/benchmark.php?test=mandelbrot&lang=all&data=u64q

[thestinger]

@TeXitoi: Using target-feature is definitely the wrong way to do it. It should be using target-cpu, not enabling individual features.

[thestinger]

Using target-cpu gives it all of the features available on that CPU (not just SSE2) and makes it use the correct cost model. The C code is compiled with -march=native so the Rust code should be using core2 as the target (or whatever the correct one is for the benchmarks game computers).

[Ryman]

Can confirm that I get the same results with -C target-cpu=core2 on an i5 (with i686-apple-darwin), thanks @thestinger!

[TeXitoi]

Maybe we can close this issue in favor of #18085 ?

[thestinger]

Sure, it covers the same thing. I altered the title a bit to include the fact that one is missing.