add patch for GCCcore 11.1.0 + 11.2.0 to fix AVX2 bug#17135
add patch for GCCcore 11.1.0 + 11.2.0 to fix AVX2 bug#17135boegel merged 1 commit intoeasybuilders:developfrom
Conversation
|
Test report by @Flamefire |
|
@boegelbot please test @ generoso |
|
@boegel: Request for testing this PR well received on login1 PR test command '
Test results coming soon (I hope)... Details- notification for comment with ID 1385929660 processed Message to humans: this is just bookkeeping information for me, |
|
Test report by @boegelbot |
|
@boegelbot please test @ jsc-zen2 |
|
@boegel: Request for testing this PR well received on jsczen2l1.int.jsc-zen2.easybuild-test.cluster PR test command '
Test results coming soon (I hope)... Details- notification for comment with ID 1386033358 processed Message to humans: this is just bookkeeping information for me, |
|
Test report by @boegelbot |
|
Test report by @Flamefire |
|
Test report by @boegel |
|
Going in, thanks @Flamefire! |
|
@boegelbot please test @ generoso |
|
@boegel: Request for testing this PR well received on login1 PR test command '
Test results coming soon (I hope)... Details- notification for comment with ID 1386934278 processed Message to humans: this is just bookkeeping information for me, |
|
Test report by @verdurin |
|
Test report by @boegelbot |
|
@boegelbot please test @ jsc-zen2 |
|
@boegel: Request for testing this PR well received on jsczen2l1.int.jsc-zen2.easybuild-test.cluster PR test command '
Test results coming soon (I hope)... Details- notification for comment with ID 1387138137 processed Message to humans: this is just bookkeeping information for me, |
|
Test report by @Flamefire |
|
Test report by @boegelbot |
(created using
eb --new-pr)This fixes a serious bug when using AVX-2 intrinsics such as done by XNNPACK, a dependency of PyTorch and TensorFlow for quantization. It yields wrong results as the order of operands in the AVX vector is wrong when using an affected intrinsic and due to aliasing issues further bugs are possible based on "undefined behavior", i.e.: "everything" is possible.
See e.g. pytorch/pytorch#92246 for an actual bug caused by this and https://stackoverflow.com/a/72837992/1930508 for the post that led me to the solution in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99754