Skip to content

Improve performance of CUDA kernels generated from manual_sparse_jacobian #4093

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
Nov 21, 2025
+17 −14
Merged

Improve performance of CUDA kernels generated from manual_sparse_jacobian #4093

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
0576df8
Improve performance of Cuda kernel by fusing and reordering operations
daverumph Nov 3, 2025
76fe001
Faster implementation of a couple of lines
daverumph Nov 11, 2025
87354a9
Change order of calculation in two places of update_jacobian
daverumph Nov 11, 2025
8e4eb69
Merge branch 'main' into dr/gpu_perf/manual_sparse_jacobian_1
daverumph Nov 19, 2025
1202522
Increment ref_counter.jl
daverumph Nov 21, 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
7 changes: 6 additions & 1 deletion reproducibility_tests/ref_counter.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
274
275

# **README**
#
Expand All @@ -20,6 +20,11 @@


#=
275
- Change order of GPU calculations for better performance, but it
results in slightly different floating point rounding. Artifacts
all look the same.
274
- Remove unused calculation of TKE exchange in mixing length
Expand Down
24 changes: 11 additions & 13 deletions src/prognostic_equations/implicit/manual_sparse_jacobian.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -626,11 +626,13 @@ function update_jacobian!(alg::ManualSparseJacobian, cache, Y, p, dtγ, t)
FT = eltype(params)
(; vertical_diffusion, smagorinsky_lilly) = p.atmos
(; ᶜp) = p.precomputed
ᶜK_u = p.scratch.ᶜtemp_scalar_4
ᶜK_h = p.scratch.ᶜtemp_scalar_6
if vertical_diffusion isa DecayWithHeightDiffusion
ᶜK_h = ᶜcompute_eddy_diffusivity_coefficient(Y.c.ρ, vertical_diffusion)
ᶜK_h .= ᶜcompute_eddy_diffusivity_coefficient(Y.c.ρ, vertical_diffusion)
ᶜK_u = ᶜK_h
elseif vertical_diffusion isa VerticalDiffusion
ᶜK_h = ᶜcompute_eddy_diffusivity_coefficient(Y.c.uₕ, ᶜp, vertical_diffusion)
ᶜK_h .= ᶜcompute_eddy_diffusivity_coefficient(Y.c.uₕ, ᶜp, vertical_diffusion)
ᶜK_u = ᶜK_h
elseif is_smagorinsky_vertical(smagorinsky_lilly)
set_smagorinsky_lilly_precomputed_quantities!(Y, p, smagorinsky_lilly)
Expand All @@ -650,22 +652,18 @@ function update_jacobian!(alg::ManualSparseJacobian, cache, Y, p, dtγ, t)
@. ᶜK_h = eddy_diffusivity(ᶜK_u, ᶜprandtl_nvec)
end


@. p.scratch.ᶜbidiagonal_adjoint_matrix_c3 =
ᶜadvdivᵥ_matrix() ⋅ DiagonalMatrixRow(ᶠinterp(ᶜρ) * ᶠinterp(ᶜK_h))
@. ᶜdiffusion_h_matrix =
p.scratch.ᶜbidiagonal_adjoint_matrix_c3 ⋅ ᶠgradᵥ_matrix()

∂ᶠρχ_dif_flux_∂ᶜχ = ᶠp_grad_matrix
@. ∂ᶠρχ_dif_flux_∂ᶜχ =
DiagonalMatrixRow(ᶠinterp(ᶜρ) * ᶠinterp(ᶜK_h)) ⋅ ᶠgradᵥ_matrix()
@. ᶜdiffusion_h_matrix = ᶜadvdivᵥ_matrix() ⋅ ∂ᶠρχ_dif_flux_∂ᶜχ
if (
MatrixFields.has_field(Y, @name(c.sgs⁰.ρatke)) ||
!isnothing(p.atmos.turbconv_model) ||
!disable_momentum_vertical_diffusion(p.atmos.vertical_diffusion)
)
@. p.scratch.ᶜbidiagonal_adjoint_matrix_c3 =
ᶜadvdivᵥ_matrix() ⋅
DiagonalMatrixRow(ᶠinterp(ᶜρ) * ᶠinterp(ᶜK_u))
@. ᶜdiffusion_u_matrix =
p.scratch.ᶜbidiagonal_adjoint_matrix_c3 ⋅ ᶠgradᵥ_matrix()
@. ∂ᶠρχ_dif_flux_∂ᶜχ =
DiagonalMatrixRow(ᶠinterp(ᶜρ) * ᶠinterp(ᶜK_u)) ⋅ ᶠgradᵥ_matrix()
@. ᶜdiffusion_u_matrix = ᶜadvdivᵥ_matrix() ⋅ ∂ᶠρχ_dif_flux_∂ᶜχ
end

∂ᶜρe_tot_err_∂ᶜρ = matrix[@name(c.ρe_tot), @name(c.ρ)]
Expand Down
Loading