Adapt to CUDAapi and CUDAdrv changes.

Author: maleadt

Manifest.toml

@@ -16,17 +16,15 @@ version = "0.2.0"
 
 [[CUDAapi]]
 deps = ["Libdl", "Logging"]
-git-tree-sha1 = "56a813440ac98a1aa64672ab460a1512552211a7"
+git-tree-sha1 = "d9614968b9a13df433870115acff20f41e7b400a"
 uuid = "3895d2a7-ec45-59b8-82bb-cfc6a382f9b3"
-version = "2.1.0"
+version = "3.0.0"
 
 [[CUDAdrv]]
 deps = ["CEnum", "CUDAapi", "Printf"]
-git-tree-sha1 = "fcb6f610289397bdadac2c0aca0acc5be60bc832"
-repo-rev = "181808ff35f9deb7360edf556d4131dc55ebeee5"
-repo-url = "https://github.com/JuliaGPU/CUDAdrv.jl.git"
+git-tree-sha1 = "01e90fa34e25776bc7c8661183d4519149ebfe59"
 uuid = "c5f51814-7f29-56b8-a69c-e4d8f6be1fde"
-version = "5.1.0"
+version = "6.0.0"
 
 [[DataStructures]]
 deps = ["InteractiveUtils", "OrderedCollections"]

Project.toml

@@ -17,8 +17,8 @@ TimerOutputs = "a759f4b9-e2f1-59dc-863e-4aeb61b1ea8f"
 [compat]
 Adapt = "0.4, 1.0"
 CEnum = "0.2"
-CUDAapi = "2.0"
-CUDAdrv = "5.0"
+CUDAapi = "3.0"
+CUDAdrv = "6.0"
 DataStructures = "0.15, 0.16, 0.17"
 LLVM = "1.2"
 TimerOutputs = "0.5"

src/CUDAnative.jl

@@ -15,6 +15,15 @@ using Libdl
 
 ## global state
 
+const toolkit_dirs = Ref{Vector{String}}()
+
+"""
+    prefix()
+
+Returns the installation prefix directories of the CUDA toolkit in use.
+"""
+prefix() = toolkit_dirs[]
+
 const toolkit_version = Ref{VersionNumber}()
 
 """
@@ -39,7 +48,6 @@ const ptx_support = Ref{Vector{VersionNumber}}()
 const libdevice = Ref{Union{String,Dict{VersionNumber,String}}}()
 const libcudadevrt = Ref{String}()
 const nvdisasm = Ref{String}()
-const ptxas = Ref{String}()
 
 
 ## source code includes
@@ -109,58 +117,54 @@ function __init__()
 
         # CUDA
 
-        toolkit_dirs = find_toolkit()
-        toolkit_version[] = find_toolkit_version(toolkit_dirs)
+        toolkit_dirs[] = find_toolkit()
+
+        let val = find_cuda_binary("nvdisasm", toolkit_dirs[])
+            val === nothing && error("Your CUDA installation does not provide the nvdisasm binary")
+            nvdisasm[] = val
+        end
+
+        toolkit_version[] = parse_toolkit_version(nvdisasm[])
         if release() < v"9"
             silent || @warn "CUDAnative.jl only supports CUDA 9.0 or higher (your toolkit provides CUDA $(release()))"
         elseif release() > CUDAdrv.release()
             silent || @warn """You are using CUDA toolkit $(release()) with a driver that only supports up to $(CUDAdrv.release()).
                                It is recommended to upgrade your driver."""
         end
 
-        llvm_support = llvm_compat(llvm_version)
-        cuda_support = cuda_compat()
-
-        target_support[] = sort(collect(llvm_support.cap ∩ cuda_support.cap))
-        isempty(target_support[]) && error("Your toolchain does not support any device capability")
-
-        ptx_support[] = sort(collect(llvm_support.ptx ∩ cuda_support.ptx))
-        isempty(ptx_support[]) && error("Your toolchain does not support any PTX ISA")
-
-        @debug("CUDAnative supports devices $(verlist(target_support[])); PTX $(verlist(ptx_support[]))")
-
-        let val = find_libdevice(target_support[], toolkit_dirs)
+        let val = find_libdevice(toolkit_dirs[])
             val === nothing && error("Your CUDA installation does not provide libdevice")
             libdevice[] = val
         end
 
-        let val = find_libcudadevrt(toolkit_dirs)
+        let val = find_libcudadevrt(toolkit_dirs[])
             val === nothing && error("Your CUDA installation does not provide libcudadevrt")
             libcudadevrt[] = val
         end
 
-        let val = find_cuda_binary("nvdisasm", toolkit_dirs)
-            val === nothing && error("Your CUDA installation does not provide the nvdisasm binary")
-            nvdisasm[] = val
-        end
-
-        let val = find_cuda_binary("ptxas", toolkit_dirs)
-            val === nothing && error("Your CUDA installation does not provide the ptxas binary")
-            ptxas[] = val
-        end
-
-        let val = find_cuda_library("nvtx", toolkit_dirs)
+        let val = find_cuda_library("nvtx", toolkit_dirs[], [v"1"])
             val === nothing && error("Your CUDA installation does not provide the NVTX library")
             NVTX.libnvtx[] = val
         end
 
-        toolkit_extras_dirs = filter(dir->isdir(joinpath(dir, "extras")), toolkit_dirs)
+        toolkit_extras_dirs = filter(dir->isdir(joinpath(dir, "extras")), toolkit_dirs[])
         cupti_dirs = map(dir->joinpath(dir, "extras", "CUPTI"), toolkit_extras_dirs)
-        let val = find_cuda_library("cupti", cupti_dirs)
+        let val = find_cuda_library("cupti", cupti_dirs, [toolkit_version[]])
             val === nothing && error("Your CUDA installation does not provide the CUPTI library")
             CUPTI.libcupti[] = val
         end
 
+        llvm_support = llvm_compat(llvm_version)
+        cuda_support = cuda_compat()
+
+        target_support[] = sort(collect(llvm_support.cap ∩ cuda_support.cap))
+        isempty(target_support[]) && error("Your toolchain does not support any device capability")
+
+        ptx_support[] = sort(collect(llvm_support.ptx ∩ cuda_support.ptx))
+        isempty(ptx_support[]) && error("Your toolchain does not support any PTX ISA")
+
+        @debug("CUDAnative supports devices $(verlist(target_support[])); PTX $(verlist(ptx_support[]))")
+
 
         ## actual initialization
 

src/compatibility.jl

@@ -6,6 +6,23 @@ const highest = v"999"
 verlist(vers) = join(map(ver->"$(ver.major).$(ver.minor)", sort(collect(vers))), ", ", " and ")
 
 
+## version range
+
+struct VersionRange
+    lower::VersionNumber
+    upper::VersionNumber
+end
+
+Base.in(v::VersionNumber, r::VersionRange) = (v >= r.lower && v <= r.upper)
+
+import Base.(:)
+(:)(a::VersionNumber, b::VersionNumber) = VersionRange(a, b)
+
+Base.intersect(v::VersionNumber, r::VersionRange) =
+    v < r.lower ? (r.lower:v) :
+    v > r.upper ? (v:r.upper) : (v:v)
+
+
 ## devices supported by the CUDA toolkit
 
 # Source:

src/exceptions.jl

@@ -27,9 +27,7 @@ function create_exceptions!(mod::CuModule)
         flag_ptr[] = reinterpret(Ptr{Cvoid}, convert(CuPtr{Cvoid}, exception_flag))
     catch err
         # modules that do not throw exceptions will not contain the indicator flag
-        if err !== CUDAdrv.ERROR_NOT_FOUND
-            rethrow()
-        end
+        err.code == CUDAdrv.ERROR_NOT_FOUND || rethrow()
     end
 
     return

test/runtests.jl

@@ -9,6 +9,7 @@ include("util.jl")
 
 @test CUDAnative.functional()
 
+CUDAnative.prefix()
 CUDAnative.version()
 
 CUDAnative.enable_timings()