Merge pull request #514 from JuliaGPU/tb/libname

Use at-runtime_ccall's ability to delay the library lookup.

Author: maleadt

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")

src/blas/libcublas.jl

@@ -16,21 +16,7 @@ import ..CuArrays.unsafe_free!
 
 import NNlib
 
-const libcudnn = 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 "cudnn64_10.dll"
-    toolkit = find_toolkit()
-    path = find_cuda_library("cudnn", toolkit)
-    if path === nothing
-        nothing
-    else
-        basename(path)
-    end
-else
-    # ccall by soname; CuArrays.__init__ will have populated Libdl.DL_LOAD_PATH
-    "libcudnn"
-end
+const libcudnn = Ref("libcudnn")
 
 # core library
 include("libcudnn_common.jl")

src/dnn/CUDNN.jl

@@ -10,20 +10,7 @@ using CUDAdrv: CUstream
 
 using CEnum
 
-const libcufft = 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 "cufft64_10.dll"
-    toolkit = find_toolkit()
-    path = find_cuda_library("cufft", toolkit)
-    if path === nothing
-        error("Could not find libcufft")
-    end
-    basename(path)
-else
-    # ccall by soname; CuArrays.__init__ will have populated Libdl.DL_LOAD_PATH
-    "libcufft"
-end
+const libcufft = Ref("libcufft")
 
 # core library
 include("libcufft_common.jl")

src/dnn/libcudnn.jl

@@ -3,53 +3,53 @@
 
 
 function cufftPlan1d(plan, nx, type, batch)
-    @check @runtime_ccall((:cufftPlan1d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftPlan1d, libcufft[]), cufftResult,
                  (Ptr{cufftHandle}, Cint, cufftType, Cint),
                  plan, nx, type, batch)
 end
 
 function cufftPlan2d(plan, nx, ny, type)
-    @check @runtime_ccall((:cufftPlan2d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftPlan2d, libcufft[]), cufftResult,
                  (Ptr{cufftHandle}, Cint, Cint, cufftType),
                  plan, nx, ny, type)
 end
 
 function cufftPlan3d(plan, nx, ny, nz, type)
-    @check @runtime_ccall((:cufftPlan3d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftPlan3d, libcufft[]), cufftResult,
                  (Ptr{cufftHandle}, Cint, Cint, Cint, cufftType),
                  plan, nx, ny, nz, type)
 end
 
 function cufftPlanMany(plan, rank, n, inembed, istride, idist, onembed, ostride, odist,
                        type, batch)
-    @check @runtime_ccall((:cufftPlanMany, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftPlanMany, libcufft[]), cufftResult,
                  (Ptr{cufftHandle}, Cint, Ptr{Cint}, Ptr{Cint}, Cint, Cint, Ptr{Cint},
                   Cint, Cint, cufftType, Cint),
                  plan, rank, n, inembed, istride, idist, onembed, ostride, odist, type,
                  batch)
 end
 
 function cufftMakePlan1d(plan, nx, type, batch, workSize)
-    @check @runtime_ccall((:cufftMakePlan1d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftMakePlan1d, libcufft[]), cufftResult,
                  (cufftHandle, Cint, cufftType, Cint, Ptr{Csize_t}),
                  plan, nx, type, batch, workSize)
 end
 
 function cufftMakePlan2d(plan, nx, ny, type, workSize)
-    @check @runtime_ccall((:cufftMakePlan2d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftMakePlan2d, libcufft[]), cufftResult,
                  (cufftHandle, Cint, Cint, cufftType, Ptr{Csize_t}),
                  plan, nx, ny, type, workSize)
 end
 
 function cufftMakePlan3d(plan, nx, ny, nz, type, workSize)
-    @check @runtime_ccall((:cufftMakePlan3d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftMakePlan3d, libcufft[]), cufftResult,
                  (cufftHandle, Cint, Cint, Cint, cufftType, Ptr{Csize_t}),
                  plan, nx, ny, nz, type, workSize)
 end
 
 function cufftMakePlanMany(plan, rank, n, inembed, istride, idist, onembed, ostride, odist,
                            type, batch, workSize)
-    @check @runtime_ccall((:cufftMakePlanMany, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftMakePlanMany, libcufft[]), cufftResult,
                  (cufftHandle, Cint, Ptr{Cint}, Ptr{Cint}, Cint, Cint, Ptr{Cint}, Cint,
                   Cint, cufftType, Cint, Ptr{Csize_t}),
                  plan, rank, n, inembed, istride, idist, onembed, ostride, odist, type,
@@ -58,7 +58,7 @@ end
 
 function cufftMakePlanMany64(plan, rank, n, inembed, istride, idist, onembed, ostride,
                              odist, type, batch, workSize)
-    @check @runtime_ccall((:cufftMakePlanMany64, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftMakePlanMany64, libcufft[]), cufftResult,
                  (cufftHandle, Cint, Ptr{Clonglong}, Ptr{Clonglong}, Clonglong, Clonglong,
                   Ptr{Clonglong}, Clonglong, Clonglong, cufftType, Clonglong, Ptr{Csize_t}),
                  plan, rank, n, inembed, istride, idist, onembed, ostride, odist, type,
@@ -67,147 +67,147 @@ end
 
 function cufftGetSizeMany64(plan, rank, n, inembed, istride, idist, onembed, ostride,
                             odist, type, batch, workSize)
-    @check @runtime_ccall((:cufftGetSizeMany64, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftGetSizeMany64, libcufft[]), cufftResult,
                  (cufftHandle, Cint, Ptr{Clonglong}, Ptr{Clonglong}, Clonglong, Clonglong,
                   Ptr{Clonglong}, Clonglong, Clonglong, cufftType, Clonglong, Ptr{Csize_t}),
                  plan, rank, n, inembed, istride, idist, onembed, ostride, odist, type,
                  batch, workSize)
 end
 
 function cufftEstimate1d(nx, type, batch, workSize)
-    @check @runtime_ccall((:cufftEstimate1d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftEstimate1d, libcufft[]), cufftResult,
                  (Cint, cufftType, Cint, Ptr{Csize_t}),
                  nx, type, batch, workSize)
 end
 
 function cufftEstimate2d(nx, ny, type, workSize)
-    @check @runtime_ccall((:cufftEstimate2d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftEstimate2d, libcufft[]), cufftResult,
                  (Cint, Cint, cufftType, Ptr{Csize_t}),
                  nx, ny, type, workSize)
 end
 
 function cufftEstimate3d(nx, ny, nz, type, workSize)
-    @check @runtime_ccall((:cufftEstimate3d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftEstimate3d, libcufft[]), cufftResult,
                  (Cint, Cint, Cint, cufftType, Ptr{Csize_t}),
                  nx, ny, nz, type, workSize)
 end
 
 function cufftEstimateMany(rank, n, inembed, istride, idist, onembed, ostride, odist, type,
                            batch, workSize)
-    @check @runtime_ccall((:cufftEstimateMany, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftEstimateMany, libcufft[]), cufftResult,
                  (Cint, Ptr{Cint}, Ptr{Cint}, Cint, Cint, Ptr{Cint}, Cint, Cint,
                   cufftType, Cint, Ptr{Csize_t}),
                  rank, n, inembed, istride, idist, onembed, ostride, odist, type, batch,
                  workSize)
 end
 
 function cufftCreate(handle)
-    @check @runtime_ccall((:cufftCreate, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftCreate, libcufft[]), cufftResult,
                  (Ptr{cufftHandle},),
                  handle)
 end
 
 function cufftGetSize1d(handle, nx, type, batch, workSize)
-    @check @runtime_ccall((:cufftGetSize1d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftGetSize1d, libcufft[]), cufftResult,
                  (cufftHandle, Cint, cufftType, Cint, Ptr{Csize_t}),
                  handle, nx, type, batch, workSize)
 end
 
 function cufftGetSize2d(handle, nx, ny, type, workSize)
-    @check @runtime_ccall((:cufftGetSize2d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftGetSize2d, libcufft[]), cufftResult,
                  (cufftHandle, Cint, Cint, cufftType, Ptr{Csize_t}),
                  handle, nx, ny, type, workSize)
 end
 
 function cufftGetSize3d(handle, nx, ny, nz, type, workSize)
-    @check @runtime_ccall((:cufftGetSize3d, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftGetSize3d, libcufft[]), cufftResult,
                  (cufftHandle, Cint, Cint, Cint, cufftType, Ptr{Csize_t}),
                  handle, nx, ny, nz, type, workSize)
 end
 
 function cufftGetSizeMany(handle, rank, n, inembed, istride, idist, onembed, ostride,
                           odist, type, batch, workArea)
-    @check @runtime_ccall((:cufftGetSizeMany, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftGetSizeMany, libcufft[]), cufftResult,
                  (cufftHandle, Cint, Ptr{Cint}, Ptr{Cint}, Cint, Cint, Ptr{Cint}, Cint,
                   Cint, cufftType, Cint, Ptr{Csize_t}),
                  handle, rank, n, inembed, istride, idist, onembed, ostride, odist, type,
                  batch, workArea)
 end
 
 function cufftGetSize(handle, workSize)
-    @check @runtime_ccall((:cufftGetSize, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftGetSize, libcufft[]), cufftResult,
                  (cufftHandle, Ptr{Csize_t}),
                  handle, workSize)
 end
 
 function cufftSetWorkArea(plan, workArea)
-    @check @runtime_ccall((:cufftSetWorkArea, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftSetWorkArea, libcufft[]), cufftResult,
                  (cufftHandle, CuPtr{Cvoid}),
                  plan, workArea)
 end
 
 function cufftSetAutoAllocation(plan, autoAllocate)
-    @check @runtime_ccall((:cufftSetAutoAllocation, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftSetAutoAllocation, libcufft[]), cufftResult,
                  (cufftHandle, Cint),
                  plan, autoAllocate)
 end
 
 function cufftExecC2C(plan, idata, odata, direction)
-    @check @runtime_ccall((:cufftExecC2C, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftExecC2C, libcufft[]), cufftResult,
                  (cufftHandle, CuPtr{cufftComplex}, CuPtr{cufftComplex}, Cint),
                  plan, idata, odata, direction)
 end
 
 function cufftExecR2C(plan, idata, odata)
-    @check @runtime_ccall((:cufftExecR2C, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftExecR2C, libcufft[]), cufftResult,
                  (cufftHandle, CuPtr{cufftReal}, CuPtr{cufftComplex}),
                  plan, idata, odata)
 end
 
 function cufftExecC2R(plan, idata, odata)
-    @check @runtime_ccall((:cufftExecC2R, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftExecC2R, libcufft[]), cufftResult,
                  (cufftHandle, CuPtr{cufftComplex}, CuPtr{cufftReal}),
                  plan, idata, odata)
 end
 
 function cufftExecZ2Z(plan, idata, odata, direction)
-    @check @runtime_ccall((:cufftExecZ2Z, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftExecZ2Z, libcufft[]), cufftResult,
                  (cufftHandle, CuPtr{cufftDoubleComplex}, CuPtr{cufftDoubleComplex}, Cint),
                  plan, idata, odata, direction)
 end
 
 function cufftExecD2Z(plan, idata, odata)
-    @check @runtime_ccall((:cufftExecD2Z, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftExecD2Z, libcufft[]), cufftResult,
                  (cufftHandle, CuPtr{cufftDoubleReal}, CuPtr{cufftDoubleComplex}),
                  plan, idata, odata)
 end
 
 function cufftExecZ2D(plan, idata, odata)
-    @check @runtime_ccall((:cufftExecZ2D, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftExecZ2D, libcufft[]), cufftResult,
                  (cufftHandle, CuPtr{cufftDoubleComplex}, CuPtr{cufftDoubleReal}),
                  plan, idata, odata)
 end
 
 function cufftSetStream(plan, stream)
-    @check @runtime_ccall((:cufftSetStream, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftSetStream, libcufft[]), cufftResult,
                  (cufftHandle, CUstream),
                  plan, stream)
 end
 
 function cufftDestroy(plan)
-    @check @runtime_ccall((:cufftDestroy, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftDestroy, libcufft[]), cufftResult,
                  (cufftHandle,),
                  plan)
 end
 
 function cufftGetVersion(version)
-    @check @runtime_ccall((:cufftGetVersion, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftGetVersion, libcufft[]), cufftResult,
                  (Ptr{Cint},),
                  version)
 end
 
 function cufftGetProperty(type, value)
-    @check @runtime_ccall((:cufftGetProperty, libcufft), cufftResult,
+    @check @runtime_ccall((:cufftGetProperty, libcufft[]), cufftResult,
                  (libraryPropertyType, Ptr{Cint}),
                  type, value)
 end

src/fft/CUFFT.jl

@@ -12,20 +12,7 @@ import CUDAnative
 
 using CEnum
 
-const libcurand = 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 "curand64_10.dll"
-    toolkit = find_toolkit()
-    path = find_cuda_library("curand", toolkit)
-    if path === nothing
-        error("Could not find libcurand")
-    end
-    basename(path)
-else
-    # ccall by soname; CuArrays.__init__ will have populated Libdl.DL_LOAD_PATH
-    "libcurand"
-end
+const libcurand = Ref("libcurand")
 
 # core library
 include("libcurand_common.jl")