Temporary Fix for FP16 -> FP8 conversion failure on -0.0

[yucai-intel]

To resolve #2219
This PR is to temporarily work around the issue where FP16's -0.0 is erroneously converted to NaN during certain fusion passes (fp16 -> fp32 -> fp8), we are currently avoiding the use of the sycl::half data type in the intermediate conversion steps to prevent the problematic fusion from occurring.

[guangyey]

One question: how do you identify this is a compiler issue, any reproducer founded or regression compiler version detected?

[CuiYifeng]

One question: how do you identify this is a compiler issue, any reproducer founded or regression compiler version detected?

@guangyey Thanks for the question. We found that this issue does not occur with the following explicit fp16->fp32->fp8 conversion:

import torch
x = torch.tensor(-0.0, dtype=torch.float16).xpu()
y = x.to(torch.float32)
z = y.to(torch.float8_e4m3fn)
print(z)

however, we will get nan in the following usage with implicit fp16->fp32 conversion:

import torch
x = torch.tensor(-0.0, dtype=torch.float16).xpu()
z = x.to(torch.float8_e4m3fn)
print(z)

The key difference between these two cases is that the conversion in the first case is submitted as two kernels, but the conversion in the second one is submitted as one kernel, where some optimizations exist in the second case. Such conjecture has been confirmed by a reproducer in local.
Furthermore, we are currently not sure whether the problem is caused by compiler or IGC, so I have updated PR description.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[CuiYifeng]

@guangyey @EikanWang This fixing has been updated. Please take a look, thanks.

[guangyey]

c10::detail::fp16_ieee_to_fp32_value(src_val.x) is functionally equivalent to the fallback path where FP8 values are first converted to FP32 on the CPU. I don't know what the root cause is, however, the workaround seems good to me.
Let's @EikanWang make the final stamp.