__mm256_srl_epi64() returns different results on LDC when -mattr=+avx2 is on #143
Labels
Comments
|
Hello, your unittest seems to pass here, what's your LDC version? (EDIT: and OS?) |
|
1.11.20 now implements I think you got trapped by an old LDC promoting your int THAT SAID, it seems newer LDC prevent such implicit conversions. As said, I don't repro your unittest, which is odd. |
|
Thanks for responding and implementing _mm256_srli_epi64()! Looking at the output more, I realized while my application was being compiled with -mattr=+avx2, intel-intrinsics was not. After changing: to and also replacing _mm256_srl_epi64() with _mm256_srli_epi64(), my unittest now passes; though it does not pass when intel-intrinsics is not built with -mattr=+avx2 This is reproducible by me on debian 12 bookworm with both the repo version of LDC v1.30.0 and a freshly compiled v1.39.0 |
|
I still can't repro on Windows or godbolt, I'm going to left it here. Probably need your LLVM version with |
|
Sorry for taking so long to reply. Best I can tell is that the fallback implementation of _mm256_srli_epi64() (when -mattr=+avx2 is not passed in) is incorrect: |
|
Your compiler is a LDC 1.30 with LLVM 14.0.6 I can only get a LDC 1.30 with LLVM 14.0.3 How do you explain that you have a LDC based upon LLVM 14.0.6? When the official Linux x86_64 build here of LDC 1.30 is LLVM 14.0.3, see here https://github.com/ldc-developers/ldc/releases/tag/v1.30.0 No luck reproducing that either in Linux or Windows. |
|
I am using the version of LDC shipped by debian 12. I will try to reproduce later tonight with the upstream LDC 1.40 release |
|
Here is me reproducing the issue with LDC 1.40: |
|
verbose dub's output with ldc 1.40: |
|
Reproduced something, thanks! Sounds related to what you said indeed.
import std.stdio;
void main()
{
import inteli.avx2intrin;
import std.stdio;
long4 start = [0xffff_ffff_0000_0000,
0xffff_ffff_0000_0000,
0xffff_ffff_0000_0000,
0xffff_ffff_0000_0000];
long4 res = _mm256_srli_epi64(start, 32);
long4 expected = [0x0000_0000_ffff_ffff,
0x0000_0000_ffff_ffff,
0x0000_0000_ffff_ffff,
0x0000_0000_ffff_ffff];
writefln!("start:\n%(0x%08x %)")(start.array);
writefln!"after shift of 32 bits\n%(0x%08x %)"(res.array);
assert(res.array == expected.array);
}To reproduce: LDC 1.24 or LDC 1.28 + LDC 1.24 or 1.28 + After LDC 1.30 the result changes :) LDC 1.30 or LDC 1.36 or LDC 1.40 + LDC 1.30 or 1.36 or LDC 1.40 + Workarounds:
Reproduced with: |
|
is a codegen issue, need reduced test case but that one necessitate two dub package it seems |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The following code fails on my machine when compiling (ldc) with -mattr=+avx2
When NOT building with -mattr=+avx2:
When building with -mattr=+avx2:
CPU information:
If _mm256_srli_epi64() was implemented I would just use that instead :)
The text was updated successfully, but these errors were encountered: