search

diku-dk / futhark-benchmarks

Various Futhark benchmark programs
Other
40 stars 7 forks source link

Gram matrix initialization #15

Open UnixJunkie opened 3 years ago

UnixJunkie commented 3 years ago

Dear futharckers, I wrote this recently: https://github.com/ocaml-bench/sandmark/blob/master/benchmarks/multicore-grammatrix/grammatrix.ml The Gram matrix is quite useful in several machine learning settings. I think it might be a nice, real-world, addition to the futhark benchmarks. Futharkly yours, F.

UnixJunkie commented 3 years ago

Some people wrote a multithread version of it: https://github.com/ocaml-bench/sandmark/blob/master/benchmarks/multicore-grammatrix/grammatrix_multicore.ml I have a parallel (multicore) version of it there: https://github.com/UnixJunkie/gram-matrix-bench/blob/master/src/gram.ml

UnixJunkie commented 3 years ago

My bench results can be seen there: https://github.com/UnixJunkie/gram-matrix-bench

athas commented 3 years ago

Is there much to gain by exploiting the symmetry, when compared to just doing a normal matrix multiplication? It seems like it would only reduce the work by a factor of two.

UnixJunkie commented 3 years ago

The matrix is diagonal symmetric. Usually, I only compute one half plus the diagonal. Also, this matrix can be quite big quite fast, because of the dataset size.

UnixJunkie commented 3 years ago

Also, probably stupid question, can we read a file in Futhark, and write to one? I have read that there is no IO in the futhark doc, so I am really wondering.

athas commented 3 years ago

No, Futhark is pure. For testing and benchmarking programs, the compiler automatically inserts IO for reading input and writing results (see here), but for actual production use, Futhark code must be invoked from a conventional programming language (discussed here).

UnixJunkie commented 3 years ago

Wow, amazing. Thanks for the pointers. I read that: "Simplicity. I think modern computers are too slow, too fragile, and too incorrect, and I believe the root cause is uncontrolled complexity. Attempts to introduce simplicity by mindlessly adding abstraction layers are ultimately counterproductive. Simplicity is best achieved by lowering feature count and flexibility." Now, I feel less lonely. :)