Merge branch 'master' of https://github.com/vargonis/CuArrays.jl

Author: Rodrigo Vargas

.gitlab-ci.yml

@@ -111,7 +111,7 @@ flux:
   tags:
     - nvidia
   script:
-    - export FLUX="$HOME/.julia/dev/Flux"
+    - export FLUX=".julia/dev/Flux"
     - julia -e 'using Pkg;
                 Pkg.develop("Flux");'
     - julia --project -e 'using Pkg;

Manifest.toml

@@ -50,12 +50,6 @@ git-tree-sha1 = "ed2c4abadf84c53d9e58510b5fc48912c2336fbb"
 uuid = "34da2185-b29b-5c13-b0c7-acf172513d20"
 version = "2.2.0"
 
-[[Crayons]]
-deps = ["Test"]
-git-tree-sha1 = "f621b8ef51fd2004c7cf157ea47f027fdeac5523"
-uuid = "a8cc5b0e-0ffa-5ad4-8c14-923d3ee1735f"
-version = "4.0.0"
-
 [[DataStructures]]
 deps = ["InteractiveUtils", "OrderedCollections"]
 git-tree-sha1 = "0809951a1774dc724da22d26e4289bbaab77809a"
@@ -176,10 +170,10 @@ deps = ["Distributed", "InteractiveUtils", "Logging", "Random"]
 uuid = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [[TimerOutputs]]
-deps = ["Crayons", "Printf", "Test", "Unicode"]
-git-tree-sha1 = "b80671c06f8f8bae08c55d67b5ce292c5ae2660c"
+deps = ["Printf"]
+git-tree-sha1 = "311765af81bbb48d7bad01fb016d9c328c6ede03"
 uuid = "a759f4b9-e2f1-59dc-863e-4aeb61b1ea8f"
-version = "0.5.0"
+version = "0.5.3"
 
 [[Tokenize]]
 git-tree-sha1 = "dfcdbbfb2d0370716c815cbd6f8a364efb6f42cf"

src/CuArrays.jl

@@ -55,9 +55,8 @@ const __initialized__ = Ref(false)
 functional() = __initialized__[]
 
 export has_cudnn, has_cutensor
-const libraries = Dict{String,Union{String,Nothing}}()
-has_cudnn() = libraries["cudnn"] !== nothing && CUDNN.libcudnn !== nothing
-has_cutensor() = libraries["cutensor"] !== nothing && CUTENSOR.libcutensor !== nothing
+has_cudnn() = Libdl.dlopen_e(CUDNN.libcudnn[]) !== C_NULL
+has_cutensor() = Libdl.dlopen_e(CUTENSOR.libcutensor[]) !== C_NULL
 
 function __init__()
     precompiling = ccall(:jl_generating_output, Cint, ()) != 0
@@ -76,18 +75,20 @@ function __init__()
         for name in ("cublas", "cusparse", "cusolver", "cufft", "curand", "cudnn", "cutensor")
             mod = getfield(CuArrays, Symbol(uppercase(name)))
             lib = Symbol("lib$name")
-            path = find_cuda_library(name, toolkit)
-            libraries[name] = path
-
-            # only push the load path if we couldn't find the library
-            if path !== nothing
-                file = basename(path)
-                handle = first(split(file,'.'))
-                if Libdl.dlopen_e(handle) == C_NULL
-                    dir = dirname(path)
-                    if !(dir in Libdl.DL_LOAD_PATH)
-                        push!(Libdl.DL_LOAD_PATH, dir)
-                    end
+            handle = getfield(mod, lib)
+
+            # on Windows, the library name is version dependent
+            if Sys.iswindows()
+                cuda = CUDAnative.version()
+                suffix = cuda >= v"10.1" ? "$(cuda.major)" : "$(cuda.major)$(cuda.minor)"
+                handle[] = "$(name)$(Sys.WORD_SIZE)_$(suffix)"
+            end
+
+            # check if we can't find the library
+            if Libdl.dlopen_e(handle[]) == C_NULL
+                path = find_cuda_library(name, toolkit)
+                if path !== nothing
+                    handle[] = path
                 end
             end
         end

src/blas/CUBLAS.jl

@@ -13,20 +13,7 @@ using LinearAlgebra
 
 using CEnum
 
-const libcublas = if Sys.iswindows()
-    # no ccall by soname, we need the filename
-    # NOTE: we discover the full path here, while only the wordsize and toolkit versions
-    #       would have been enough to construct "cublas64_10.dll"
-    toolkit = find_toolkit()
-    path = find_cuda_library("cublas", toolkit)
-    if path === nothing
-        error("Could not find libcublas")
-    end
-    const libcublas = basename(path)
-else
-    # ccall by soname; CuArrays.__init__ will have populated Libdl.DL_LOAD_PATH
-    const libcublas = "libcublas"
-end
+const libcublas = Ref("libcublas")
 
 # core library
 include("libcublas_common.jl")