Add warning for mGGA (#28)

.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.3.13
+current_version = 0.3.14
 commit = True
 tag = False
 

.gitlab-ci.yml

@@ -18,7 +18,8 @@ julia/1.8-n2:
     - if: $CI_PIPELINE_SOURCE == "external_pull_request_event"
       when: manual
   script:
-    - module load lang/JuliaHPC/1.8.3-foss-2022a-CUDA-11.7.0
+    # - module load lang/JuliaHPC/1.8.3-foss-2022a-CUDA-11.7.0
+    - module load lang/Julia/1.8.3-linux-x86_64
     - julia --color=yes --project=. -e '
         using Pkg;
         Pkg.test(; coverage=true, test_args=["gpu"])

Project.toml

@@ -1,14 +1,16 @@
 name = "Libxc"
 uuid = "66e17ffc-8502-11e9-23b5-c9248d0eb96d"
 authors = ["Jason Eu <[email protected]>", "Michael F. Herbst <[email protected]>"]
-version = "0.3.13"
+version = "0.3.14"
 
 [deps]
+CUDA_Runtime_jll = "76a88914-d11a-5bdc-97e0-2f5a05c973a2"
 Libxc_GPU_jll = "25af9330-9b41-55d4-a324-1a83c0a0a1ac"
 Libxc_jll = "a56a6d9d-ad03-58af-ab61-878bf78270d6"
 Requires = "ae029012-a4dd-5104-9daa-d747884805df"
 
 [compat]
+CUDA_Runtime_jll = "0.2"
 Libxc_GPU_jll = "6"
 Libxc_jll = "6"
 Requires = "1"

src/Libxc.jl

@@ -12,7 +12,7 @@ const libxc_version = VersionNumber(XC_VERSION)
 const libxc_doi = unsafe_string(Libxc.xc_reference_doi())
 
 """Is the CUDA version of libxc available on this platform"""
-has_cuda() = isdefined(Libxc_GPU_jll, :libxc)
+has_cuda() = Libxc_GPU_jll.is_available()
 
 """Return the list of available libxc functionals as strings"""
 function available_functionals()
@@ -35,7 +35,7 @@ export needs_laplacian
 
 function __init__()
     @require CUDA="052768ef-5323-5732-b1bb-66c8b64840ba" begin
-        @static if isdefined(Libxc_GPU_jll, :libxc)
+        @static if Libxc_GPU_jll.is_available()
             include("evaluate_gpu.jl")
         end
     end

src/evaluate_gpu.jl

@@ -143,19 +143,24 @@ function evaluate!(func::Functional, ::Union{Val{:mgga},Val{:hyb_mgga}}, rho::Cu
                    v4lapltau3::OptCuArray=CU_NULL,
                    v4tau4::OptCuArray=CU_NULL)
     np = Int(length(rho) / func.spin_dimensions.rho)
+    @warn "meta-GGAs on GPU seem to be broken at least with Libxc 6.1.0"
 
     pointer = allocate_gpufunctional(func)
-    @ccall libxc_gpu.xc_gga(
-        pointer::Ptr{xc_func_type}, np::Csize_t, rho::CuPtr{Cdouble}, sigma::CuPtr{Cdouble},
+    @ccall libxc_gpu.xc_mgga(
+        pointer::Ptr{xc_func_type}, np::Csize_t,
+        rho::CuPtr{Cdouble}, sigma::CuPtr{Cdouble},
         lapl::CuPtr{Cdouble}, tau::CuPtr{Cdouble},
+        #
         zk::CuPtr{Cdouble}, vrho::CuPtr{Cdouble}, vsigma::CuPtr{Cdouble},
         vlapl::CuPtr{Cdouble}, vtau::CuPtr{Cdouble},
+        #
         v2rho2::CuPtr{Cdouble}, v2rhosigma::CuPtr{Cdouble},
         v2rholapl::CuPtr{Cdouble}, v2rhotau::CuPtr{Cdouble},
         v2sigma2::CuPtr{Cdouble}, v2sigmalapl::CuPtr{Cdouble},
         v2sigmatau::CuPtr{Cdouble}, v2lapl2::CuPtr{Cdouble},
         v2lapltau::CuPtr{Cdouble}, v2tau2::CuPtr{Cdouble},
         v3rho3::CuPtr{Cdouble}, v3rho2sigma::CuPtr{Cdouble},
+        #
         v3rho2lapl::CuPtr{Cdouble}, v3rho2tau::CuPtr{Cdouble},
         v3rhosigma2::CuPtr{Cdouble}, v3rhosigmalapl::CuPtr{Cdouble},
         v3rhosigmatau::CuPtr{Cdouble}, v3rholapl2::CuPtr{Cdouble},
@@ -166,24 +171,24 @@ function evaluate!(func::Functional, ::Union{Val{:mgga},Val{:hyb_mgga}}, rho::Cu
         v3lapl3::CuPtr{Cdouble}, v3lapl2tau::CuPtr{Cdouble},
         v3lapltau2::CuPtr{Cdouble}, v3tau3::CuPtr{Cdouble},
         v4rho4::CuPtr{Cdouble}, v4rho3sigma::CuPtr{Cdouble},
-        v4rho3lapl::CuPtr{Cdouble}, v4rho3tau::CuPtr{Cdouble},
-        v4rho2sigma2::CuPtr{Cdouble}, v4rho2sigmalapl::CuPtr{Cdouble},
-        v4rho2sigmatau::CuPtr{Cdouble}, v4rho2lapl2::CuPtr{Cdouble},
-        v4rho2lapltau::CuPtr{Cdouble}, v4rho2tau2::CuPtr{Cdouble},
-        v4rhosigma3::CuPtr{Cdouble}, v4rhosigma2lapl::CuPtr{Cdouble},
-        v4rhosigma2tau::CuPtr{Cdouble}, v4rhosigmalapl2::CuPtr{Cdouble},
-        v4rhosigmalapltau::CuPtr{Cdouble}, v4rhosigmatau2::CuPtr{Cdouble},
-        v4rholapl3::CuPtr{Cdouble}, v4rholapl2tau::CuPtr{Cdouble},
-        v4rholapltau2::CuPtr{Cdouble}, v4rhotau3::CuPtr{Cdouble},
-        v4sigma4::CuPtr{Cdouble}, v4sigma3lapl::CuPtr{Cdouble},
-        v4sigma3tau::CuPtr{Cdouble}, v4sigma2lapl2::CuPtr{Cdouble},
-        v4sigma2lapltau::CuPtr{Cdouble}, v4sigma2tau2::CuPtr{Cdouble},
-        v4sigmalapl3::CuPtr{Cdouble}, v4sigmalapl2tau::CuPtr{Cdouble},
-        v4sigmalapltau2::CuPtr{Cdouble}, v4sigmatau3::CuPtr{Cdouble},
-        v4lapl4::CuPtr{Cdouble}, v4lapl3tau::CuPtr{Cdouble},
-        v4lapl2tau2::CuPtr{Cdouble}, v4lapltau3::CuPtr{Cdouble},
-        v4tau4::CuPtr{Cdouble}
+        v4rho3lapl::CuPtr{Cdouble},
+        #
+        v4rho3tau::CuPtr{Cdouble}, v4rho2sigma2::CuPtr{Cdouble},
+        v4rho2sigmalapl::CuPtr{Cdouble}, v4rho2sigmatau::CuPtr{Cdouble},
+        v4rho2lapl2::CuPtr{Cdouble}, v4rho2lapltau::CuPtr{Cdouble},
+        v4rho2tau2::CuPtr{Cdouble}, v4rhosigma3::CuPtr{Cdouble},
+        v4rhosigma2lapl::CuPtr{Cdouble}, v4rhosigma2tau::CuPtr{Cdouble},
+        v4rhosigmalapl2::CuPtr{Cdouble}, v4rhosigmalapltau::CuPtr{Cdouble},
+        v4rhosigmatau2::CuPtr{Cdouble}, v4rholapl3::CuPtr{Cdouble},
+        v4rholapl2tau::CuPtr{Cdouble}, v4rholapltau2::CuPtr{Cdouble},
+        v4rhotau3::CuPtr{Cdouble}, v4sigma4::CuPtr{Cdouble},
+        v4sigma3lapl::CuPtr{Cdouble}, v4sigma3tau::CuPtr{Cdouble},
+        v4sigma2lapl2::CuPtr{Cdouble}, v4sigma2lapltau::CuPtr{Cdouble},
+        v4sigma2tau2::CuPtr{Cdouble}, v4sigmalapl3::CuPtr{Cdouble},
+        v4sigmalapl2tau::CuPtr{Cdouble}, v4sigmalapltau2::CuPtr{Cdouble},
+        v4sigmatau3::CuPtr{Cdouble}, v4lapl4::CuPtr{Cdouble},
+        v4lapl3tau::CuPtr{Cdouble}, v4lapl2tau2::CuPtr{Cdouble},
+        v4lapltau3::CuPtr{Cdouble}, v4tau4::CuPtr{Cdouble},
     )::Cvoid
-
     deallocate_gpufunctional(pointer)
 end

test/gpu.jl

@@ -3,7 +3,7 @@ using Libxc
 using CUDA
 
 @testset "Properties" begin
-    @test CUDA.has_cuda()
+    @test Libxc.has_cuda()
 end
 
 @testset "Compare LDA on CPU/GPU" begin
@@ -33,8 +33,31 @@ end
         sigma_d = convert(CuArray, sigma)
         res_d   = evaluate(Functional(sym; n_spin); rho=rho_d, sigma=sigma_d)
 
-        @test maximum(abs,   res.zk   - Array(res_d.zk))     < 1e-12
+        @test maximum(abs, res.zk     - Array(res_d.zk))     < 1e-12
         @test maximum(abs, res.vrho   - Array(res_d.vrho))   < 1e-12
         @test maximum(abs, res.vsigma - Array(res_d.vsigma)) < 1e-12
     end
 end
+
+
+@testset "Compare mGGA without Laplacian" begin
+    n_p = 12
+
+    for n_spin in (1, 2), sym in (:mgga_x_tpss, )
+        rho   = abs.(randn(n_spin,    n_p))
+        sigma = abs.(randn(2n_spin-1, n_p))
+        tau   = abs.(randn(n_spin,    n_p))
+        res   = evaluate(Functional(sym; n_spin); rho, sigma, tau)
+
+        rho_d   = convert(CuArray, rho)
+        sigma_d = convert(CuArray, sigma)
+        tau_d   = convert(CuArray, tau)
+        res_d   = evaluate(Functional(sym; n_spin);
+                           rho=rho_d, sigma=sigma_d, tau=tau_d)
+
+        @test_broken maximum(abs, res.zk     - Array(res_d.zk))     < 1e-12
+        @test_broken maximum(abs, res.vrho   - Array(res_d.vrho))   < 1e-12
+        @test_broken maximum(abs, res.vsigma - Array(res_d.vsigma)) < 1e-12
+        @test_broken maximum(abs, res.vtau   - Array(res_d.vtau))   < 1e-12
+    end
+end