Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Math] Fix NaN from powf() domain error. #2067

Open
wants to merge 2 commits into
base: main
Choose a base branch
from

Conversation

stolk
Copy link

@stolk stolk commented Sep 30, 2024

This change prevents calling powf(x,y) with negative x.

The SIMD versions using ssePower() already seem to be resistent to negative pixel values, but the scalar version was not.

This would cause a SIGFPE on apps that have floating point exceptions enabled.

FIXES: #2066

Copy link

linux-foundation-easycla bot commented Sep 30, 2024

CLA Signed

The committers listed above are authorized under a signed CLA.

Copy link
Collaborator

@doug-walker doug-walker left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks Bram, good catch. What do you think about moving the std::pow call down into the ternary operator, to avoid adding the extra std::max call?

In addition, I think this would fix the unit test errors. (You're clamping the value but the ternary expression needs the unclamped value.)

@stolk
Copy link
Author

stolk commented Sep 30, 2024

Thanks Bram, good catch. What do you think about moving the std::pow call down into the ternary operator, to avoid adding the extra std::max call?

Hi Doug,

I advise against the ternary operator...

I checked, and the compilers (gcc, clang) actually produce more efficient code with std::max than with ?:

See this godbolt output.

@stolk
Copy link
Author

stolk commented Sep 30, 2024

...not authorized...

I expected having a CLA in place for OpenImageIO with ASWF, would transfer over to OpenColorIO?

@stolk
Copy link
Author

stolk commented Sep 30, 2024

In addition, I think this would fix the unit test errors. (You're clamping the value but the ternary expression needs the unclamped value.)

Oops, did I break the unit test? I'll revisit.

@doug-walker
Copy link
Collaborator

I was not proposing to replace the max with a new ternary. I was suggesting moving the pow call down into the existing ternary. As I mentioned, your unit tests are failing, so some action is needed.

The CLAs for all ASWF projects are managed through the same account but need to be signed separately for each project.

@stolk
Copy link
Author

stolk commented Sep 30, 2024

I was not proposing to replace the max with a new ternary. I was suggesting moving the pow call down into the existing ternary. As I mentioned, your unit tests are failing, so some action is needed.

Thanks Doug, I will redo this PR.

Strangely, the unmodified main branch fails the unit test as well. I have to find out what is going on here.

$ make test
Running tests...
Test project /home/stolk/src/OpenColorIO/build
    Start 1: test_utils_exec
1/6 Test #1: test_utils_exec ..................   Passed    0.01 sec
    Start 2: test_cpu
2/6 Test #2: test_cpu .........................***Failed    5.78 sec
    Start 3: test_cpu_no_accel
3/6 Test #3: test_cpu_no_accel ................***Failed    5.41 sec
    Start 4: test_cpu_avx512
4/6 Test #4: test_cpu_avx512 ..................***Failed    5.99 sec
    Start 5: test_gpu
5/6 Test #5: test_gpu .........................***Failed    8.83 sec
    Start 6: test_python
6/6 Test #6: test_python ......................   Passed    1.38 sec

33% tests passed, 4 tests failed out of 6

Does the main branch HEAD pass the unit tests for you? Thanks.

@doug-walker
Copy link
Collaborator

All the tests should pass, except for the GPU tests on some mac platforms currently.

This change prevents calling powf(x,y) with negative x.

The SIMD versions using ssePower() already seem to be resistent to
negative pixel values, but the scalar version was not.

This would cause a SIGFPE on apps that have floating point exceptions
enabled.

FIXES: AcademySoftwareFoundation#2066

Signed-off-by: Bram Stolk <[email protected]>
@stolk
Copy link
Author

stolk commented Sep 30, 2024

I have changed this PR to be less invasive, and only correct the value passed to powf() and not the entire pixel value as used in the rest of the function.

Note: I cannot get the unit test to pass on my machine, even when using HEAD of main. Pushed the PR to see if CI will pass.

@KevinJW
Copy link
Contributor

KevinJW commented Oct 2, 2024

just to note, micro optimisation difference comes from the testing vs return order, if you flip the ternary around you can get the same code ... https://godbolt.org/z/Wzafa3nsb

@stolk
Copy link
Author

stolk commented Oct 10, 2024

Can a maintainer please approve the CI run, and then merge? Thanks.

Copy link
Collaborator

@doug-walker doug-walker left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi Bram, what do you think about changing this:

out[0] = pixel[0]<=red[3] ? pixel[0] * red[4] : data[0];

to this:

out[0] = pixel[0]<=red[3] ? pixel[0] * red[4] :
std::pow(pixel[0] * red[0] + red[1], red[2]);

That way it avoids the extra branch.

@stolk
Copy link
Author

stolk commented Oct 11, 2024

Hi Bram, what do you think about changing this:

out[0] = pixel[0]<=red[3] ? pixel[0] * red[4] : data[0];

to this:

out[0] = pixel[0]<=red[3] ? pixel[0] * red[4] : std::pow(pixel[0] * red[0] + red[1], red[2]);

That way it avoids the extra branch.

No, that does not address the issue.

We should not be feeding negative numbers to the base of a powf().

My change makes sure of that, your version does not.

@doug-walker
Copy link
Collaborator

Hi Bram,

Is the reason that you don't like my proposal because you are worried that pixel[0] * red[0] + red[1] will be negative? In practice, that is never the case. In theory, someone may be able to set parameter values that would place the breakpoint below zero, but if that is the edge case you are worried about, it would be better dealt with in the GammaOpData.cpp:GammaOpData::validateParameters function rather than inside the performance critical pixel loop.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

Floating Point Domain Error triggered by IdentifyBuiltinColorSpace()
3 participants