Skip to content

Add support for fp16 comparison native operators #8010

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.

Sign up for GitHub

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

Merged
sklam merged 16 commits into from
Jul 27, 2022
Merged

Add support for fp16 comparison native operators #8010

Changes from 1 commit
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
c33108a
Add support for fp16 comparison operators
testhound Apr 27, 2022
c353f64
Correctly handle diffeerent types when declaring fp16 comparison oper…
testhound May 12, 2022
0752ccd
Merge branch 'main' into testhound/fp16-comparison-operators
testhound May 12, 2022
d9abed6
Rename class to clarify it's purpose
testhound May 16, 2022
5fe0392
Allow safe conversions from int8 to float16
testhound May 25, 2022
9ead75c
Allow safe conversions for fp16 conversions
testhound May 25, 2022
4a2006d
Add test cases for mixed type comparisions with fp16
testhound May 25, 2022
77b745e
Merge branch 'main' into testhound/fp16-comparison-operators
testhound May 25, 2022
5663396
Remove superflous comment and document logic in _genfp16_comparison_o…
testhound Jul 5, 2022
35fd82f
Add setUp to seed RNG, remove unnecessary assignments, modify code to…
testhound Jul 5, 2022
0b6d045
Merge branch 'main' into testhound/fp16-comparison-operators
testhound Jul 5, 2022
6e02ce8
Fix test cases
testhound Jul 7, 2022
9911814
Add comment explaining why AbstractTemplate and ConcreteTemplate are …
testhound Jul 7, 2022
da6016a
Add failing test case for multiple comparison test
testhound Jul 7, 2022
400ac81
Merge branch 'main' into testhound/fp16-comparison-operators
testhound Jul 7, 2022
8be1d0c
Update numba/cuda/cudadecl.py
testhound Jul 21, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Update numba/cuda/cudadecl.py 
Use fleshed out comment description suggested by @gmarkall

Co-authored-by: Graham Markall <535640+gmarkall@users.noreply.github.com>
  • Loading branch information
@testhound @gmarkall
testhound and gmarkall authored Jul 21, 2022
commit 8be1d0c9c2d4290880e3605c625c9c53495572e6
16 changes: 12 additions & 4 deletions numba/cuda/cudadecl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -379,10 +379,18 @@ class Cuda_fp16_cmp(ConcreteTemplate):
return Cuda_fp16_cmp


# Until issue 7863 is resolved, if multiple concrete templates are utilized
# function resolution will pick the wrong implementation and errors will occur.
# We avoid this by using ConcreteTempalte for the intrinisics and
# AbstractTemplate
# If multiple ConcreteTemplates provide typing for a single function, then
# function resolution will pick the first compatible typing it finds even if it
# involves inserting a cast that would be considered undesirable (in this
# specific case, float16s could be cast to float32s for comparisons).
#
# To work around this, we instead use an AbstractTemplate that implements
# exactly the casting logic that we desire. The AbstractTemplate gets
# considered in preference to ConcreteTemplates during typing.
#
# This is tracked as Issue #7863 (https://github.com/numba/numba/issues/7863) -
# once this is resolved it should be possible to replace this AbstractTemplate
# with a ConcreteTemplate to simplify the logic.
def _genfp16_comparison_operator(l_key):
@register_global(l_key)
class Cuda_fp16_operator_cmp(AbstractTemplate):
Expand Down