Defer initializing deps to actual run time.

Author: maleadt

.gitlab-ci.yml

@@ -5,7 +5,6 @@ image: ubuntu:bionic
 
 variables:
   JULIA_DEBUG: 'CuArrays'
-  JULIA_CUDA_VERBOSE: 'true'
   CI_APT_INSTALL: 'libgomp1'
   NVIDIA_VISIBLE_DEVICES: 'all'
   NVIDIA_DRIVER_CAPABILITIES: 'compute,utility'

Manifest.toml

@@ -34,14 +34,16 @@ version = "3.1.0"
 
 [[CUDAdrv]]
 deps = ["CEnum", "CUDAapi", "Printf"]
-git-tree-sha1 = "01e90fa34e25776bc7c8661183d4519149ebfe59"
+git-tree-sha1 = "cdf9a88aae8f05df1b78b049ffafbdc6d41a650e"
+repo-rev = "ab4d8e3f449b71104c57f827c67567c34c04a5ff"
+repo-url = "https://github.com/JuliaGPU/CUDAdrv.jl.git"
 uuid = "c5f51814-7f29-56b8-a69c-e4d8f6be1fde"
 version = "6.0.0"
 
 [[CUDAnative]]
 deps = ["Adapt", "BinaryProvider", "CEnum", "CUDAapi", "CUDAdrv", "DataStructures", "InteractiveUtils", "LLVM", "Libdl", "MacroTools", "Pkg", "Printf", "TimerOutputs"]
-git-tree-sha1 = "5beaa8e6a8fb65003b7b2555a73e7666468b22f6"
-repo-rev = "46d5331"
+git-tree-sha1 = "3dd3ef011ea2a8f5b0e915588f3c2a3cfd68aadd"
+repo-rev = "11146e91cb3bd1f86c32b936abc25dea10c7e6cc"
 repo-url = "https://github.com/JuliaGPU/CUDAnative.jl.git"
 uuid = "be33ccc6-a3ff-5ff2-a52e-74243cff1e17"
 version = "2.10.2"
@@ -79,6 +81,7 @@ uuid = "929cbde3-209d-540e-8aea-75f648917ca0"
 version = "1.3.3"
 
 [[LibGit2]]
+deps = ["Printf"]
 uuid = "76f85450-5226-5b5a-8eaa-529ad045b433"
 
 [[Libdl]]
@@ -114,7 +117,7 @@ uuid = "bac558e1-5e72-5ebc-8fee-abe8a469f55d"
 version = "1.1.0"
 
 [[Pkg]]
-deps = ["Dates", "LibGit2", "Libdl", "Logging", "Markdown", "Printf", "REPL", "Random", "SHA", "Test", "UUIDs"]
+deps = ["Dates", "LibGit2", "Libdl", "Logging", "Markdown", "Printf", "REPL", "Random", "SHA", "UUIDs"]
 uuid = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 
 [[Printf]]

src/CuArrays.jl

@@ -49,43 +49,15 @@ include("deprecated.jl")
 
 ## initialization
 
-const __initialized__ = Ref(false)
-functional() = __initialized__[]
+function __init__()
+    # package integrations
+    @require ForwardDiff="f6369f11-7733-5829-9624-2563aa707210" include("forwarddiff.jl")
 
-export has_cudnn, has_cutensor
-has_cudnn() = Libdl.dlopen_e(CUDNN.libcudnn[]) !== C_NULL
-has_cutensor() = Libdl.dlopen_e(CUTENSOR.libcutensor[]) !== C_NULL
+    __init_memory__()
 
-function __init__()
-    precompiling = ccall(:jl_generating_output, Cint, ()) != 0
-    silent = parse(Bool, get(ENV, "JULIA_CUDA_SILENT", "false")) || precompiling
-    verbose = parse(Bool, get(ENV, "JULIA_CUDA_VERBOSE", "false"))
-
-    # if any dependent GPU package failed, expect it to have logged an error and bail out
-    if !CUDAdrv.functional() || !CUDAnative.functional()
-        verbose && @warn "CuArrays.jl did not initialize because CUDAdrv.jl or CUDAnative.jl failed to"
-        return
-    end
-
-    try
-        __init_bindeps__(silent=silent, verbose=verbose)
-
-        # package integrations
-        @require ForwardDiff="f6369f11-7733-5829-9624-2563aa707210" include("forwarddiff.jl")
-
-        __init_memory__()
-
-        __initialized__[] = true
-    catch ex
-        # don't actually fail to keep the package loadable
-        if !silent
-            if verbose
-                @error "CuArrays.jl failed to initialize" exception=(ex, catch_backtrace())
-            else
-                @info "CuArrays.jl failed to initialize and will be unavailable (set JULIA_CUDA_SILENT or JULIA_CUDA_VERBOSE to silence or expand this message)"
-            end
-        end
-    end
+    # NOTE: we only perform minimal initialization here that does not require CUDA or a GPU.
+    #       most of the actual initialization is deferred to run time:
+    #       see bindeps.jl for initialization of CUDA binary dependencies.
 end
 
 end # module