[PATCH] D99741: [RISCV][Clang] Add some RVV Floating-Point intrinsic functions. (vfclass, vfmerge, vfrec7, vfrsqrt7, vfsqrt)
Craig Topper via Phabricator via cfe-commits
cfe-commits at lists.llvm.org
Sat May 1 14:53:31 PDT 2021
craig.topper added a comment.
In D99741#2730989 <https://reviews.llvm.org/D99741#2730989>, @thakis wrote:
>> ! In D99741#2721669 <https://reviews.llvm.org/D99741#2721669>, @craig.topper wrote:
>> We're preparing a patch to remove to stop testing both rv32 and rv64 on every test. That should reduce the time by half. What is an acceptable number?
>
> You tell me. I'd say well below 1%?
>
> I wanted to prepare a graph that shows check-clang time over the last year – it stayed basically constant and the jumped by 20% (now 10% now that half the tests are disabled). It'd look dramatic, but I haven't had time to make a graph. I'm sure you can imagine it though :) (Also, I feel the burden of proof shouldn't be on people pointing out regressions.)
>
> @rnk suggested putting these tests behind some kind of `REQUIRES: expensive-checks` thing that's off by default, instead of reverting everything, until perf is sorted out. That sounds like a great idea to me – wdyt?
I just uploaded D101700 <https://reviews.llvm.org/D101700> which takes lit on `clang/test/CodeGen/RISCV/rvv*` from ~8 seconds to ~5.5 seconds as measured on a 48 core server I use at work. Not sure how that translates to a less capable machine with less cores, slower disk, etc.
For comparison, here are some times for running lit on subsets of tests on this machine before the patch. Mostly looked at clang/test/CodeGen as I was curious to see how close we are to other targets.
clang/test/ - 68.75 seconds
clang/test/CodeGen - 24.59 seconds
clang/test/CodeGen/X86 - 2.68 seconds
clang/test/CodeGen/aarch64* - 12.33 seconds
clang/test/CodeGen/arm* - 2.04 seconds
We're also considering D100529 <https://reviews.llvm.org/D100529> to further reduce the RISCV test time. Another option we've talked about is concatenating some of the tests into a single file to cut down the number of times we parse riscv_vector.h.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D99741/new/
https://reviews.llvm.org/D99741
More information about the cfe-commits
mailing list