Construct c10::Half from float16_t on ARMv8

[malfet]

By hiding float32 constructors and exposing float16 ones. This allows compiler do implicit conversions as needed, and in safe cases optimize out unneeded upcasts to fp32, see example below

#include <arm_neon.h>

#ifndef __ARM_FEATURE_FP16_SCALAR_ARITHMETIC
#error Ieeee
#endif

float16_t sum1(float16_t x, float16_t y) {
    return x + y;
}

float16_t sum2(float16_t x, float16_t y) {
    return static_cast<float>(x) + static_cast<float>(y);
}

both sum variants are compiled to scalar fp16 add, if build for the platform that supports fp16 arithmetic

sum1(half, half):                            // @sum1(half, half)
        fadd    h0, h0, h1
        ret
sum2(half, half):                            // @sum2(half, half)
        fadd    h0, h0, h1
        ret

Fixes build error in some aarch64 configurations after #119483 which are defined as supporting FP16 but don't define _Float16.

cc @snadampal

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/120425

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (5 Unrelated Failures)

As of commit 734f0b3 with merge base 65627cf ():

FLAKY - The following jobs failed but were likely due to flakiness present on trunk:

• linux-binary-manywheel / manywheel-py3_10-cpu-cxx11-abi-build / build (gh)
  ../c10/util/C++17.h:13:2: error: #error "You're trying to build PyTorch with a too old version of GCC. We need GCC 9 or later."
• linux-binary-manywheel / manywheel-py3_11-cpu-cxx11-abi-build / build (gh)
  ../c10/util/C++17.h:13:2: error: #error "You're trying to build PyTorch with a too old version of GCC. We need GCC 9 or later."
• linux-binary-manywheel / manywheel-py3_12-cpu-cxx11-abi-build / build (gh)
  ../c10/util/C++17.h:13:2: error: #error "You're trying to build PyTorch with a too old version of GCC. We need GCC 9 or later."
• linux-binary-manywheel / manywheel-py3_8-cpu-cxx11-abi-build / build (gh)
  ../c10/util/C++17.h:13:2: error: #error "You're trying to build PyTorch with a too old version of GCC. We need GCC 9 or later."
• linux-binary-manywheel / manywheel-py3_9-cpu-cxx11-abi-build / build (gh)
  ../c10/util/C++17.h:13:2: error: #error "You're trying to build PyTorch with a too old version of GCC. We need GCC 9 or later."

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[mikekgfb]

LGTM!

[atalman]

@malfet Since #119483 was reverted should we merge both PRs, test and land them together ?

[malfet]

@malfet Since #119483 was reverted should we merge both PRs, test and land them together ?

If this can be landed, it potentially achieves the same results as the other PR, but with fewer lines of code

[malfet]

@snadampal FYI

[malfet]

@pytorchbot merge -i

[snadampal]

Looks good to me!

If I understand correctly, this is an extension to the fp16<->fp32 acceleration feature added as part of this PR. now it's reusing the same via Half operator. isn't it?

I'm wondering what the major use cases for fp16 (half) datatype kernels are.

[pytorchmergebot]

Merge started

Your change will be merged while ignoring the following 5 checks: linux-binary-manywheel / manywheel-py3_12-cpu-cxx11-abi-build / build, linux-binary-manywheel / manywheel-py3_11-cpu-cxx11-abi-build / build, linux-binary-manywheel / manywheel-py3_8-cpu-cxx11-abi-build / build, linux-binary-manywheel / manywheel-py3_9-cpu-cxx11-abi-build / build, linux-binary-manywheel / manywheel-py3_10-cpu-cxx11-abi-build / build

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here