Merge pull request #322 from jpsamaroo/jps/amdgpu-hip

Updates for HIP-based AMDGPU

.buildkite/runtests.yml

@@ -4,6 +4,7 @@ steps:
       setup:
         version:
           - "1"
+          - "1.9"
     env:
       GROUP: CUDA
     plugins:
@@ -31,7 +32,7 @@ steps:
       setup:
         version:
           # Known issues on 1.8
-          - "1.9-nightly"
+          - "1"
     env:
       GROUP: AMDGPU
     plugins:
@@ -59,7 +60,7 @@ steps:
       setup:
         version:
           - "1"
-          - "1.9-nightly"
+          - "1.9"
     env:
       GROUP: oneAPI
     plugins:
@@ -69,8 +70,11 @@ steps:
       - |
         julia --project=test -e '
           import Pkg
-          Pkg.add(; name = "oneAPI")'
-        rm test/Manifest.toml
+          Pkg.add(; name="oneAPI")
+
+          println("+++ :julia: Building support library")
+          include(joinpath(dirname(dirname(Base.find_package("oneAPI"))), "deps", "build_ci.jl"))
+          Pkg.activate()'
         julia --project -e '
           import Pkg
           println("+++ :julia: Running tests")
@@ -86,8 +90,7 @@ steps:
     matrix:
       setup:
         version:
-          - "1"
-          - "1.9-nightly"
+          - "1.9"
     env:
       GROUP: Metal
     plugins:

Project.toml

@@ -38,7 +38,7 @@ MetalExt = ["Metal"]
 oneAPIExt = ["oneAPI"]
 
 [compat]
-AMDGPU = "0.4.9"
+AMDGPU = "0.5, 0.6, 0.7, 0.8"
 Adapt = "3, 4"
 CUDA = "4.1.0, 5"
 ChainRulesCore = "1"

ext/AMDGPUExt.jl

@@ -5,6 +5,9 @@ import DiffEqGPU
 using .AMDGPU
 import .AMDGPU: ROCBackend
 
+function DiffEqGPU.EnsembleGPUArray(cpu_offload::Float64)
+    DiffEqGPU.EnsembleGPUArray(ROCBackend(), cpu_offload)
+end
 DiffEqGPU.maxthreads(::ROCBackend) = 256
 DiffEqGPU.maybe_prefer_blocks(::ROCBackend) = ROCBackend()
 

src/ensemblegpukernel/callbacks.jl

@@ -47,7 +47,7 @@ struct GPUContinuousCallback{F1, F2, F3, F4, F5, F6, T, T2, T3, I, R} <:
             reltol::T2,
             repeat_nudge::T3) where {F1, F2, F3, F4, F5, F6, T, T2,
             T3, I, R,
-        }
+    }
         if save_positions != (false, false)
             error("Callback `save_positions` are incompatible with kernel-based GPU ODE solvers due requiring static sizing. Please ensure `save_positions = (false,false)` is set in all callback definitions used with such solvers.")
         end

src/ensemblegpukernel/integrators/integrator_utils.jl

@@ -54,7 +54,7 @@ end
         IIP,
         S,
         T,
-    }
+}
     integrator.retcode = retcode
 end
 
@@ -156,7 +156,7 @@ end
         S,
         T,
         T1,
-    }
+}
     # Can get rid of an allocation here with a function
     # get_tmp_arr(integrator.cache) which gives a pointer to some
     # cache array which can be modified.
@@ -185,7 +185,7 @@ end
         S,
         T,
         T1,
-    }
+}
     _change_t_via_interpolation!(integrator, t, modify_save_endpoint)
 end
 
@@ -322,7 +322,7 @@ end
         IIP,
         S,
         T,
-    }
+}
     return nothing
 end
 
@@ -333,7 +333,7 @@ end
         },
         callback,
         abst) where {AlgType <: GPUODEAlgorithm, IIP, S, T
-    }
+}
     if abst == integrator.t
         tmp = integrator.u
     elseif abst == integrator.tprev

src/ensemblegpukernel/integrators/nonstiff/interpolants.jl

@@ -8,7 +8,7 @@
         IIP,
         S,
         T,
-    }
+}
     y₁ = integ.u
     k1 = integ.k1
     k2 = integ.k2

src/ensemblegpukernel/integrators/stiff/types.jl

@@ -3,13 +3,13 @@
         IIP,
         S,
         T,
-    })(t) where {
+})(t) where {
         AlgType <:
         GPUODEAlgorithm,
         IIP,
         S,
         T,
-    }
+}
     Θ = (t - integrator.tprev) / integrator.dt
     _ode_interpolant(Θ, integrator.dt, integrator.uprev, integrator)
 end

test/distributed_multi_gpu.jl

@@ -9,7 +9,6 @@ addprocs(2)
         du[2] = u[1] * (p[2] - u[3]) - u[2]
         du[3] = u[1] * u[2] - p[3] * u[3]
     end
-    CUDA.allowscalar(false)
     u0 = Float32[1.0; 0.0; 0.0]
     tspan = (0.0f0, 100.0f0)
     p = (10.0f0, 28.0f0, 8 / 3.0f0)

test/ensemblegpuarray.jl

@@ -74,15 +74,18 @@ monteprob_jac = EnsembleProblem(prob_jac, prob_func = prob_func)
 @time solve(monteprob_jac, Rodas5(), EnsembleCPUArray(), dt = 0.1,
     trajectories = 10,
     saveat = 1.0f0)
-@time solve(monteprob_jac, Rodas5(), EnsembleGPUArray(backend), dt = 0.1,
-    trajectories = 10,
-    saveat = 1.0f0)
 @time solve(monteprob_jac, TRBDF2(), EnsembleCPUArray(), dt = 0.1,
     trajectories = 10,
     saveat = 1.0f0)
-@time solve(monteprob_jac, TRBDF2(), EnsembleGPUArray(backend), dt = 0.1,
-    trajectories = 10,
-    saveat = 1.0f0)
+
+if GROUP == "CUDA"
+    @time solve(monteprob_jac, Rodas5(), EnsembleGPUArray(backend), dt = 0.1,
+        trajectories = 10,
+        saveat = 1.0f0)
+    @time solve(monteprob_jac, TRBDF2(), EnsembleGPUArray(backend), dt = 0.1,
+        trajectories = 10,
+        saveat = 1.0f0)
+end
 
 @info "Callbacks"
 
@@ -183,18 +186,19 @@ sol = solve(rober_prob, Rodas5(), abstol = 1.0f-8, reltol = 1.0f-8)
 sol = solve(rober_prob, TRBDF2(), abstol = 1.0f-4, reltol = 1.0f-1)
 rober_monteprob = EnsembleProblem(rober_prob, prob_func = prob_func)
 
-# TODO: Does not work with Linearsolve.jl v1.35.0 https://github.com/SciML/DiffEqGPU.jl/pull/229
+if GROUP == "CUDA"
+    @time sol = solve(rober_monteprob, Rodas5(),
+        EnsembleGPUArray(backend), trajectories = 10,
+        saveat = 1.0f0,
+        abstol = 1.0f-8,
+        reltol = 1.0f-8)
+    @time sol = solve(rober_monteprob, TRBDF2(),
+        EnsembleGPUArray(backend), trajectories = 10,
+        saveat = 1.0f0,
+        abstol = 1.0f-4,
+        reltol = 1.0f-1)
+end
 
-@time sol = solve(rober_monteprob, Rodas5(),
-    EnsembleGPUArray(backend), trajectories = 10,
-    saveat = 1.0f0,
-    abstol = 1.0f-8,
-    reltol = 1.0f-8)
-@time sol = solve(rober_monteprob, TRBDF2(),
-    EnsembleGPUArray(backend), trajectories = 10,
-    saveat = 1.0f0,
-    abstol = 1.0f-4,
-    reltol = 1.0f-1)
 @time sol = solve(rober_monteprob, TRBDF2(), EnsembleThreads(),
     trajectories = 10,
     abstol = 1e-4, reltol = 1e-1, saveat = 1.0f0)
@@ -241,5 +245,8 @@ monteprob = EnsembleProblem(prob_jac,
 sol = solve(monteprob, Tsit5(), EnsembleGPUArray(backend, 0.0), trajectories = 10,
     adaptive = false, dt = 0.01f0, save_everystep = false)
 
-sol = solve(monteprob, Rosenbrock23(), EnsembleGPUArray(backend, 0.0), trajectories = 10,
-    adaptive = false, dt = 0.01f0, save_everystep = false)
+if GROUP == "CUDA"
+    sol = solve(monteprob, Rosenbrock23(), EnsembleGPUArray(backend, 0.0),
+        trajectories = 10,
+        adaptive = false, dt = 0.01f0, save_everystep = false)
+end

test/ensemblegpuarray_oop.jl

@@ -34,9 +34,12 @@ prob_func = (prob, i, repeat) -> remake(prob, p = rand(Float32, 3) .* p)
 monteprob = EnsembleProblem(prob, prob_func = prob_func, safetycopy = false)
 @time sol = solve(monteprob, Tsit5(), EnsembleGPUArray(backend), trajectories = 10_000,
     saveat = 1.0f0)
-@time sol = solve(monteprob, Rosenbrock23(), EnsembleGPUArray(backend),
-    trajectories = 10_000,
-    saveat = 1.0f0)
-@time sol = solve(monteprob, TRBDF2(), EnsembleGPUArray(backend),
-    trajectories = 10_000,
-    saveat = 1.0f0)
+
+if GROUP == "CUDA"
+    @time sol = solve(monteprob, Rosenbrock23(), EnsembleGPUArray(backend),
+        trajectories = 10_000,
+        saveat = 1.0f0)
+    @time sol = solve(monteprob, TRBDF2(), EnsembleGPUArray(backend),
+        trajectories = 10_000,
+        saveat = 1.0f0)
+end

test/lower_level_api.jl

@@ -1,4 +1,4 @@
-using DiffEqGPU, StaticArrays, CUDA, Adapt, OrdinaryDiffEq
+using DiffEqGPU, StaticArrays, Adapt, OrdinaryDiffEq
 
 include("utils.jl")
 

test/runtests.jl

@@ -79,7 +79,8 @@ if GROUP in SUPPORTS_DOUBLE_PRECISION
     end
 end
 
-if GROUP == "CUDA"
+# Callbacks currently error on v1.10
+if GROUP == "CUDA" && VERSION <= v"1.9"
     # Causes dynamic function invocation
     @time @testset "GPU Kernelized Non Stiff ODE ContinuousCallback" begin
         include("gpu_kernel_de/gpu_ode_continuous_callbacks.jl")