[LLVM] Emit fp16/fp32 builtins directly into target module #12877
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Contributor
Contributor
|
@kparzysz-quic it appears one of the CI unit tests is failing: Are you able to reproduce this yourself from the LLVM versin used in CI? |
Contributor
Author
I'm pretty sure it was due to a section name that was invalid for MachO. I limited the section specification to apply to ELF only. |
For conversions between `_Float16` and `float`, LLVM uses runtime functions `__extendhfsf2` and `__truncsfhf2`. On X86 up until version 14, LLVM used `uint16_t` for representing `_Float16`. Starting with LLVM 15, half- precision values can be passed in XMM registers (i.e. as floating-point). This happens when the compilation target has SSE2 enabled (either directly, or by enabling a feature that implies SSE2). Because the names of the conversion functions remain unchanged, it is impossible for TVM to provide them in the runtime, and have them work in both cases. To solve this issue, emit these functions directly into the target module after detecting whether or not to use floating-point ABI. To allow the linker to remove potential duplicates (or if they are unused), they are weak and reside in a separate section.
Contributor
Author
|
😎 |
Member
|
Thanks @kparzysz-quic, I'll verify tomorrow. |
masahi
approved these changes
Sep 27, 2022
Member
masahi
left a comment
masahi
left a comment
There was a problem hiding this comment.
I've verified that this fixes the correctness issue I observed when using fp16 compute on LLVM 16.
supersat
added a commit
to supersat/tvm
that referenced
this pull request
Oct 20, 2022
…pache#12877)" This reverts commit f64e933.
xinetzone
pushed a commit
to daobook/tvm
that referenced
this pull request
Nov 25, 2022
) For conversions between `_Float16` and `float`, LLVM uses runtime functions `__extendhfsf2` and `__truncsfhf2`. On X86 up until version 14, LLVM used `uint16_t` for representing `_Float16`. Starting with LLVM 15, half- precision values can be passed in XMM registers (i.e. as floating-point). This happens when the compilation target has SSE2 enabled (either directly, or by enabling a feature that implies SSE2). Because the names of the conversion functions remain unchanged, it is impossible for TVM to provide them in the runtime, and have them work in both cases. To solve this issue, emit these functions directly into the target module after detecting whether or not to use floating-point ABI. To allow the linker to remove potential duplicates (or if they are unused), they are weak and reside in a separate section.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
For conversions between
_Float16andfloat, LLVM uses runtime functions__extendhfsf2and__truncsfhf2. On X86 up until version 14, LLVM useduint16_tfor representing_Float16. Starting with LLVM 15, half- precision values can be passed in XMM registers (i.e. as floating-point). This happens when the compilation target has SSE2 enabled (either directly, or by enabling a feature that implies SSE2).Because the names of the conversion functions remain unchanged, it is impossible for TVM to provide them in the runtime, and have them work in both cases. To solve this issue, emit these functions directly into the target module after detecting whether or not to use floating-point ABI. To allow the linker to remove potential duplicates (or if they are unused), they are weak and reside in a separate section.