Thanks for your pull request. It looks like this may be your first contribution to a Google open source project (if not, look below for help). Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

📝 Please visit https://cla.developers.google.com/ to sign.

Once you've signed (or fixed any issues), please reply here with @googlebot I signed it! and we'll verify it.

What to do if you already signed the CLA

Individual signers

• It's possible we don't have your GitHub username or you're using a different email address on your commit. Check your existing CLA data and verify that your email is set on your git commits.

Corporate signers

• Your company has a Point of Contact who decides which employees are authorized to participate. Ask your POC to be added to the group of authorized contributors. If you don't know who your Point of Contact is, direct the Google project maintainer to go/cla#troubleshoot (Public version).
• The email used to register you as an authorized contributor must be the email used for the Git commit. Check your existing CLA data and verify that your email is set on your git commits.
• The email used to register you as an authorized contributor must also be attached to your GitHub account.

ℹ️ Googlers: Go here for more info.

@googlebot I signed it!

CLAs look good, thanks!

ℹ️ Googlers: Go here for more info.

@N-Dekker: Could you also make this change in Eigen? The relevant file is https://bitbucket.org/eigen/eigen/src/default/Eigen/src/Core/arch/Default/Half.h.

@N-Dekker: Could you also make this change in Eigen? The relevant file is https://bitbucket.org/eigen/eigen/src/default/Eigen/src/Core/arch/Default/Half.h.

Thanks for the suggestion, @jaingaurav I never contributed to Eigen before, but I could give it a try. Specifically for their bfloat16 implementation, I think that would be this union: https://gitlab.com/libeigen/eigen/-/blob/386d809bde475c65b7940f290efe80e6a05878c4/Eigen/src/Core/arch/Default/BFloat16.h#L315 Right? Bfloat16 is my main interest for now, as I'm maintaining another implementation myself! https://github.com/biovault/biovault_bfloat16

Unfortunately, it appears that Eigen does not have a convenient bit_cast function template like absl::bit_cast. So I'd either have to propose bit_cast to Eigen as well, or fix their bfloat16 by memcpy-ing the bits from float to unsigned int, inside their float_to_bfloat16_rtne.

Anyway, it would certainly help if this pull request would make it into the master branch of TensorFlow 😃 I'll have another look tomorrow!

@N-Dekker: Yeah I'd recommend updating both Half.h & BFloat16.h.

We're trying to merge your changes but there seem to be some BUILD file issues. I think you need to include the absl dependency else we are getting errors like: 'absl/base/casts.h': No such file or directory

We're trying to merge your changes but there seem to be some BUILD file issues. I think you need to include the absl dependency else we are getting errors like: 'absl/base/casts.h': No such file or directory

Thank you @jaingaurav, but I'm sorry, I'm not really familiar with the build system! Would it be sufficient to just add "@com_google_absl//absl/base" to tensorflow/core/lib/bfloat16/BUILD ? I guess somewhere here, right?
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/lib/bfloat16/BUILD#L18

Yep I believe that should work

@N-Dekker: I believe there is a ubuntu sanity check failure due to the ordering of the build deps. Could you please address that?

I believe there is a ubuntu sanity check failure due to the ordering of the build deps. Could you please address that?

@jaingaurav: Thanks, I see now in https://source.cloud.google.com/results/invocations/c3882af9-15f4-4c40-b79a-37a233c1d006/targets/%2F%2Ftensorflow%2Ftools%2Fci_build:gen_ci_sanity_out/log

FAIL: buildifier found errors and/or warnings in above BUILD files.
buildifier suggested the following changes:
tensorflow/core/lib/bfloat16/BUILD:
18d17
<         "@com_google_absl//absl/base",
20a20
>         "@com_google_absl//absl/base",
exit status 1
Please fix manually or run buildifier <file> to auto-fix.

I guess that means that it suggests moving the line down, from 18 to 20, right? Do you have a clue why? I'm asking because other BUILD files do have "@com_google_absl//absl/base" in front of their "//tensorflow/core..." dependencies, for example: https://github.com/tensorflow/tensorflow/blob/v2.3.0-rc0/tensorflow/compiler/aot/BUILD#L43

Anyway, I'll give it a try.

@jaingaurav Still build failures, unfortunately! I guess because of the "absl/base" dependency as well. Would you suggest me to simply avoid using Abseil for this pull request? (Instead of absl::bit_cast<uint32_t>(v), the member function could directly call std::memcpy(&u, &v, sizeof(float)), of course.)

I'm asking you also because I have another pull request in mind that may or may not use bit_cast

@N-Dekker: Some of the failures seem unrelated. However, given that we're unlikely to use absl in Eigen an alternate fix without absl would be preferred.

Apologies @N-Dekker but we're actively reworking this code to use the Eigen implementation. As a result, we'll address this casting behavior when performing that work.

Apologies @N-Dekker but we're actively reworking this code to use the Eigen implementation. As a result, we'll address this casting behavior when performing that work.

No problem @jaingaurav, , thank you for the information! Would you still suggest me to make a merge request on this issue for https://gitlab.com/libeigen/eigen ?

I'm asking also because I was about to prepare another possible improvement: I believe the performance of conversion from an integer type to bfloat could be improved significantly.

@N-Dekker: I think I'd hold off just a bit until @rmlarsen completes the work. He said he might be able to fix this issue when doing that work.

If the performance improvement is isolated, maybe you can start proposing that in eigen right away?

FYI I just submitted a bfloat32 merge request to Eigen, albeit on a different subject: "Allow implicit conversion from bfloat16 to float and double", https://gitlab.com/libeigen/eigen/-/merge_requests/163

@jaingaurav FYI, The eigen::bfloat16 merge request that corresponds to this tensorflow::bfloat16 pull request has just been approved and merged onto the Eigen master branch by your Google colleague Rasmus Munk Larsen (@rmlarsen) 😃
"Avoid undefined behavior by union type punning in float_to_bfloat16_rtne"
https://gitlab.com/libeigen/eigen/-/merge_requests/164
https://gitlab.com/libeigen/eigen/-/commit/b11f817bcff04276f3024d6780f56a137968b81a

@jaingaurav

If the performance improvement is isolated, maybe you can start proposing that in eigen right away?

FYI, the eigen::bfloat16 performance improvement merge request that I submitted is currently being reviewed: "Faster conversion from integer types to bfloat16", https://gitlab.com/libeigen/eigen/-/merge_requests/166