Skip to content

Couple of minor fixes #3151

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
maleadt merged 2 commits into from
May 22, 2026
+15 −7
Merged

Couple of minor fixes #3151

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
451467a
CUPTI: Guard against zero throughput.
maleadt May 22, 2026
87f6ea0
Tests: Robustly handle LLVM 22's directed rounding emission.
maleadt May 22, 2026
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
9 changes: 7 additions & 2 deletions CUDATools/src/profile.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -934,8 +934,13 @@ function Base.show(io::IO, results::ProfileResults)
println(io, "\nDevice-side activity: GPU was busy for $(format_time(device_time)) ($(format_percentage(device_ratio)) of the trace)")
end

# add memory throughput information
device = merge(device, (; throughput=device.size ./ device.time))
# add memory throughput information. CUPTI's timestamp resolution
# can round very short events down to 0 ns, so guard against the
# resulting Inf throughput (which would later trip up format_bytes).
throughput = map(device.size, device.time) do s, t
(s === missing || !isfinite(t) || t == 0) ? missing : s / t
end
device = merge(device, (; throughput))

if isempty(device.id)
println(io, "\nNo device-side activity was recorded.")
Expand Down
13 changes: 8 additions & 5 deletions test/core/device/intrinsics/math.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -120,18 +120,21 @@ using SpecialFunctions
end
end

# NVPTX has no sub.<rnd> intrinsic; sub_<rnd>(x,y) reuses add_<rnd>(x,-y).
# For non-rn modes LLVM keeps the rounded add; for rn (the default) it
# may fold back to a plain `sub`.
# NVPTX has no `llvm.nvvm.sub.<rnd>` intrinsic, so sub_<rnd>(x,y) is
# implemented as add_<rnd>(x,-y). PTX itself does accept rounding
# modifiers on `sub`, so the backend may fold back to a real
# `sub.<rnd>.<suffix>` (LLVM 22) or keep the rounded add (older LLVM).
# For `rn` the suffix may be elided entirely.
for rnd in (:rn, :rz, :rm, :rp)
f = getfield(CUDA, Symbol(:sub_, rnd))
for (T, suffix) in ((Float32, "f32"), (Float64, "f64"))
kernel = (out, x, y) -> (out[] = f(x, y); nothing)
buf = CuArray{T}(undef, 1)
ptx = sprint(io->(@device_code_ptx io=io @cuda launch=false kernel(buf, T(1), T(1))))
accepted = rnd === :rn ?
("add.rn.$(suffix)", "add.$(suffix)", "sub.$(suffix)") :
("add.$(rnd).$(suffix)",)
("add.rn.$(suffix)", "add.$(suffix)",
"sub.rn.$(suffix)", "sub.$(suffix)") :
("add.$(rnd).$(suffix)", "sub.$(rnd).$(suffix)")
@test any(s -> occursin(s, ptx), accepted)
end
end
Expand Down