add more examples + add calling nim from python #15
Conversation
|
@Vindaar @HugoGranstrom Some improvement over speeding Python with Nim :) LMK what you think and how we can build better example (like maybe some non-dummy examples on random data ? ) |
… & fastest loop. Perf goes from 90 seconds to 500/600 ms
…cinim into clonk/improve-npy-array-interop
examples/numpyarrays/examply.nims
Outdated
| --b:cpp | ||
| --outdir:bin | ||
| --out:examply.so | ||
| --cc:gcc | ||
| --mm:arc |
There was a problem hiding this comment.
arc is equivalent to orc without cyclic type but both should work.
c++ because I tested calling some openmp loop from python and had some issue with it using C backend and gcc-14. g++ seemed to work fine so I just rolled with it.
Also found out clang and openmp do not like each other (on OSX at least).
|
Looks good to me aside from my two minor comments! |
|
Looks good to me 😄 Crazy that Nim is sooo much faster 🤯 |
|
I updated the perf, with better python perf as well. I added strides to get indexing syntax but this makes iterating over 1D a little bit slower than with rax buffer access. |
examples/numpyarrays/examply.nim
Outdated
| for i in 0||(x.shape[0]-1): | ||
| for j in 0||(x.shape[1]-1): | ||
| result[i, j] = doStuff x[i, j] |
There was a problem hiding this comment.
I don't think this compiles correctly, or does it? Last time I tried that, I ended up writing my own loop fusion macro https://github.com/Vindaar/llm.nim/blob/master/fuse_loops.nim.
There was a problem hiding this comment.
Just remembered that this was of course in the context of using an OpenMP collapse statement as well.
There was a problem hiding this comment.
And on the C backend in general || does not seem to compile at the moment due to goto statements. Only compiling with --exceptions:quirky makes it work. Maybe there's a different exception approach that also compiles, but well.
No description provided.