freenode/#sbcl - IRC Chatlog
Search
18:23:11
flip214
In case anyone wants to take a look, https://github.com/niofis/raybench has a benchmark where SBCL is 10 times slower than GCC -O3 (the double float version).
18:34:15
stassats
if that's the quality of the lisp code what can be said about the whole benchmark?
18:39:11
stassats
take (vector float) and change to (vector single-float) and you'll see a massive speed up
19:10:46
karlosz
flip214: type declarations with safety gives you non generic arithmetic with type checks. type declarations plus safety 0 just remove those type checks
19:12:10
karlosz
but once it knows that the type is what it is, it uses native machine instructions instead of generic arithemetic
19:12:29
flip214
hmmm, I still have checks for thread.profile-counter and ENABLE-ALLOC-COUNTER in my disassembly.... but ISTR that this is because of the flags I compiled SBCL with, right?
19:13:33
flip214
still, the original reference source will have them included as well, so it's still fair to compare
19:14:38
flip214
12.6% in RND-DOME... that I can't do anything about, because it's randomness filtering
20:11:57
karlosz
fiveop: did it work? you should be able to get enough to have swank up from quicklisp
20:18:00
fiveop
I'm running the following fedora version btw: 5.0.0-0.rc2.git0.1.0.riscv64.fc30.riscv64
20:28:23
karlosz
how long does it take for you to build stuff, btw? i couldnt figure out how to make qemu any faster
20:29:29
fiveop
Total build time is about less than half an hour (I didn't pay close attention so far)
20:32:57
fiveop
So make-target-contrib.sh helpfully cleans out all fasl files, but leaves the test-passed.test-report files, so that install.sh chokes trying to install sb-bsd-sockets right away