Skip to content

[libspirv][ptx-nvidiacl] Change __clc__group_scratch size to 32 x i128 #18431

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
merged 5 commits into from
May 26, 2025
Merged

[libspirv][ptx-nvidiacl] Change __clc__group_scratch size to 32 x i128 #18431

Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
0f5efc3
[libspirv][ptx-nvidiacl] Change __clc__group_scratch size to 32 x i128
wenju-he May 13, 2025
cf9ec4a
opaque pointer, add global variables for all sizes
wenju-he May 20, 2025
8898ee5
undef -> poison
wenju-he May 20, 2025
8c02604
remove gep
wenju-he May 22, 2025
0693f5e
i32 addrspace(3)* -> ptr addrspace(3)
wenju-he May 25, 2025
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
24 changes: 12 additions & 12 deletions libclc/libspirv/lib/ptx-nvidiacl/group/collectives_helpers.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,60 +2,60 @@
; 128 bits per warp is sufficient for all fundamental data types and complex
; Reducing storage for small data types or increasing it for user-defined types
; will likely require an additional pass to track group algorithm usage
@__clc__group_scratch = internal addrspace(3) global [128 x i64] undef, align 1
@__clc__group_scratch = internal addrspace(3) global [32 x i128] undef, align 1

define i8 addrspace(3)* @__clc__get_group_scratch_bool() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We could/should probably rewrite this to use opaque pointers while we're here. If we do, almost all of this can go away. You could just return ptr addrspace(3) @__clc__group_scratch for every overload, I think?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We could/should probably rewrite this to use opaque pointers while we're here.

done

If we do, almost all of this can go away. You could just return ptr addrspace(3) @__clc__group_scratch for every overload, I think?

I added more global variables for different sizes. It should resolve the comment Reducing storage for small data types or increasing it for user-defined types will likely require an additional pass to track group algorithm usage on the top of the file.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks. I take it the scratch memory isn't mean to be shared between the different types? If so we couldn''t have separate globals in this way.

%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to i8 addrspace(3)*
ret i8 addrspace(3)* %cast
}

define i8 addrspace(3)* @__clc__get_group_scratch_char() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to i8 addrspace(3)*
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If we don't switch to opaque pointers, this needs fixing. Having bitcast i64 is incorrect as %ptr would be i128 addrspace(3)*?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done

ret i8 addrspace(3)* %cast
}

define i16 addrspace(3)* @__clc__get_group_scratch_short() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to i16 addrspace(3)*
ret i16 addrspace(3)* %cast
}

define i32 addrspace(3)* @__clc__get_group_scratch_int() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to i32 addrspace(3)*
ret i32 addrspace(3)* %cast
}

define i64 addrspace(3)* @__clc__get_group_scratch_long() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to i64 addrspace(3)*
ret i64 addrspace(3)* %cast
}

define half addrspace(3)* @__clc__get_group_scratch_half() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to half addrspace(3)*
ret half addrspace(3)* %cast
}

define float addrspace(3)* @__clc__get_group_scratch_float() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to float addrspace(3)*
ret float addrspace(3)* %cast
}

define double addrspace(3)* @__clc__get_group_scratch_double() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to double addrspace(3)*
ret double addrspace(3)* %cast
}
Expand All @@ -77,21 +77,21 @@ entry:

define %complex_half addrspace(3)* @__clc__get_group_scratch_complex_half() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to %complex_half addrspace(3)*
ret %complex_half addrspace(3)* %cast
}

define %complex_float addrspace(3)* @__clc__get_group_scratch_complex_float() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to %complex_float addrspace(3)*
ret %complex_float addrspace(3)* %cast
}

define %complex_double addrspace(3)* @__clc__get_group_scratch_complex_double() nounwind alwaysinline {
entry:
%ptr = getelementptr inbounds [128 x i64], [128 x i64] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%ptr = getelementptr inbounds [32 x i128], [32 x i128] addrspace(3)* @__clc__group_scratch, i64 0, i64 0
%cast = bitcast i64 addrspace(3)* %ptr to %complex_double addrspace(3)*
ret %complex_double addrspace(3)* %cast
}
Loading