Merge pull request #2062 from JuliaGPU/tb/gpucompiler

Adapt to GPUCompiler#master.

Author: maleadt

.buildkite/pipeline.yml

@@ -107,27 +107,25 @@ steps:
             !build.pull_request.draft
         timeout_in_minutes: 120
         commands: |
-          julia -e '
+          julia --project -e '
             using Pkg
-            Pkg.activate(joinpath(pwd(), "lib", lowercase("{{matrix.package}}")))
 
             println("--- :julia: Instantiating project")
             withenv("JULIA_PKG_PRECOMPILE_AUTO" => 0) do
-              pkgs = [PackageSpec(path=pwd())]
+              Pkg.instantiate()
+
+              pkgs = [PackageSpec(path=joinpath(pwd(), "lib", lowercase("{{matrix.package}}")))]
               if "{{matrix.package}}" == "cuTensorNet"
                 # cuTensorNet depends on a development version of cuTENSOR
                 push!(pkgs, PackageSpec(path=joinpath(pwd(), "lib", "cutensor")))
               end
               Pkg.develop(pkgs)
-              Pkg.instantiate()
 
-              Pkg.add("CUDA_Runtime_jll")
-              write(joinpath(pwd(), "lib", lowercase("{{matrix.package}}"), "LocalPreferences.toml"),
-                    "[CUDA_Runtime_jll]\nversion = \"{{matrix.cuda}}\"")
+              write("LocalPreferences.toml", "[CUDA_Runtime_jll]\nversion = \"{{matrix.cuda}}\"")
             end
 
             println("+++ :julia: Running tests")
-            Pkg.test(coverage=true)'
+            Pkg.test("{{matrix.package}}"; coverage=true)'
 
   - group: ":telescope: Downstream"
     depends_on: "cuda"
@@ -142,25 +140,21 @@ steps:
                 - lib
                 - examples
         command: |
-          julia -e '
+          julia --project -e '
             using Pkg
 
             println("--- :julia: Instantiating project")
             withenv("JULIA_PKG_PRECOMPILE_AUTO" => 0) do
-              Pkg.develop("NNlibCUDA")
-              Pkg.activate("NNlibCUDA")
-
-              Pkg.develop([PackageSpec(path=pwd()),
-                           PackageSpec(path=joinpath(pwd(), "lib", "cudnn"))])
               Pkg.instantiate()
 
-              Pkg.add("CUDA_Runtime_jll")
-              write(joinpath(Pkg.devdir(), "NNlibCUDA", "LocalPreferences.toml"),
-                    "[CUDA_Runtime_jll]\nversion = \"12.0\"")
+              Pkg.develop("NNlibCUDA")
+              Pkg.develop(path=joinpath(pwd(), "lib", "cudnn"))
+
+              write("LocalPreferences.toml", "[CUDA_Runtime_jll]\nversion = \"12.0\"")
             end
 
             println("+++ :julia: Running tests")
-            Pkg.test(coverage=true)'
+            Pkg.test("NNlibCUDA"; coverage=true)'
         agents:
           queue: "juliagpu"
           cuda: "*"
@@ -292,10 +286,8 @@ steps:
             ENV["CODESPEED_ENVIRONMENT"] = ENV["BUILDKITE_AGENT_NAME"]
 
             println("--- :julia: Instantiating project")
-            Pkg.resolve()
             Pkg.instantiate()
             Pkg.activate("perf")
-            Pkg.resolve()
             Pkg.instantiate()
             push!(LOAD_PATH, @__DIR__)
 

Manifest.toml

@@ -2,7 +2,7 @@
 
 julia_version = "1.8.5"
 manifest_format = "2.0"
-project_hash = "e8fd324aba09b72a771344d3910c4e1284ae62a0"
+project_hash = "c38f3e404ea342fb740cb16012407cbac90360fd"
 
 [[deps.AbstractFFTs]]
 deps = ["ChainRulesCore", "LinearAlgebra", "Test"]
@@ -170,9 +170,11 @@ version = "0.1.5"
 
 [[deps.GPUCompiler]]
 deps = ["ExprTools", "InteractiveUtils", "LLVM", "Libdl", "Logging", "Scratch", "TimerOutputs", "UUIDs"]
-git-tree-sha1 = "8de395b1243771bbb79ac832ec96c7def7a4586f"
+git-tree-sha1 = "2bec2c336b65f1ca50c4fcb2933712503db8eb2d"
+repo-rev = "master"
+repo-url = "https://github.com/JuliaGPU/GPUCompiler.jl.git"
 uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
-version = "0.22.0"
+version = "0.23.0"
 
 [[deps.InlineStrings]]
 deps = ["Parsers"]

Project.toml

@@ -31,10 +31,15 @@ RandomNumbers = "e6cf234a-135c-5ec9-84dd-332b85af5143"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
 Requires = "ae029012-a4dd-5104-9daa-d747884805df"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
-SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 UnsafeAtomicsLLVM = "d80eeb9a-aca5-4d75-85e5-170c8b632249"
 
+[weakdeps]
+SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
+
+[extensions]
+SpecialFunctionsExt = "SpecialFunctions"
+
 [compat]
 AbstractFFTs = "0.4, 0.5, 1.0"
 Adapt = "3.3"
@@ -47,7 +52,7 @@ Crayons = "4"
 DataFrames = "1"
 ExprTools = "0.1"
 GPUArrays = "8.6"
-GPUCompiler = "0.22"
+GPUCompiler = "0.23"
 KernelAbstractions = "0.9.2"
 LLVM = "6"
 NVTX = "0.3.2"
@@ -60,3 +65,6 @@ Requires = "0.5, 1.0"
 SpecialFunctions = "1.3, 2"
 UnsafeAtomicsLLVM = "0.1"
 julia = "1.8"
+
+[extras]
+SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"

src/device/intrinsics/special_math.jl renamed to ext/SpecialFunctionsExt.jl

@@ -1,5 +1,12 @@
 # math functionality corresponding to SpecialFunctions.jl
 
+module SpecialFunctionsExt
+
+using CUDA
+using CUDA: @device_override
+
+isdefined(Base, :get_extension) ? (using SpecialFunctions) : (using ..SpecialFunctions)
+
 
 ## error
 
@@ -46,3 +53,6 @@
 
 @device_override SpecialFunctions.bessely(n::Int32, x::Float64) = ccall("extern __nv_yn", llvmcall, Cdouble, (Int32, Cdouble), n, x)
 @device_override SpecialFunctions.bessely(n::Int32, x::Float32) = ccall("extern __nv_ynf", llvmcall, Cfloat, (Int32, Cfloat), n, x)
+
+
+end

lib/cudnn/test/runtests.jl

@@ -1,15 +1,5 @@
 using Test
 
-# work around JuliaLang/Pkg.jl#2500
-if VERSION < v"1.8"
-    test_project = first(Base.load_path())
-    preferences_file = joinpath(dirname(@__DIR__), "LocalPreferences.toml")
-    test_preferences_file = joinpath(dirname(test_project), "LocalPreferences.toml")
-    if isfile(preferences_file) && !isfile(test_preferences_file)
-        cp(preferences_file, test_preferences_file)
-    end
-end
-
 using CUDA
 @info "CUDA information:\n" * sprint(io->CUDA.versioninfo(io))
 

lib/cusolver/dense.jl

@@ -22,7 +22,7 @@ for (bname, fname,elty) in ((:cusolverDnSpotrf_bufferSize, :cusolverDnSpotrf, :F
     @eval begin
         function potrf!(uplo::Char,
                         A::StridedCuMatrix{$elty})
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             n = checksquare(A)
             lda = max(1, stride(A, 2))
 
@@ -55,7 +55,7 @@ for (fname,elty) in ((:cusolverDnSpotrs, :Float32),
         function potrs!(uplo::Char,
                         A::StridedCuMatrix{$elty},
                         B::StridedCuVecOrMat{$elty})
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             n = checksquare(A)
             if size(B, 1) != n
                 throw(DimensionMismatch("first dimension of B, $(size(B,1)), must match second dimension of A, $n"))
@@ -83,7 +83,7 @@ for (bname, fname,elty) in ((:cusolverDnSpotri_bufferSize, :cusolverDnSpotri, :F
     @eval begin
         function potri!(uplo::Char,
                         A::StridedCuMatrix{$elty})
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             n = checksquare(A)
             lda = max(1, stride(A, 2))
 
@@ -177,7 +177,7 @@ for (bname, fname,elty) in ((:cusolverDnSsytrf_bufferSize, :cusolverDnSsytrf, :F
     @eval begin
         function sytrf!(uplo::Char,
                         A::StridedCuMatrix{$elty})
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             n = checksquare(A)
             lda = max(1, stride(A, 2))
 
@@ -502,7 +502,7 @@ for (jname, bname, fname, elty, relty) in ((:syevd!, :cusolverDnSsyevd_bufferSiz
         function $jname(jobz::Char,
                         uplo::Char,
                         A::StridedCuMatrix{$elty})
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             n       = checksquare(A)
             lda     = max(1, stride(A, 2))
             W       = CuArray{$relty}(undef, n)
@@ -542,7 +542,7 @@ for (jname, bname, fname, elty, relty) in ((:sygvd!, :cusolverDnSsygvd_bufferSiz
                         uplo::Char,
                         A::StridedCuMatrix{$elty},
                         B::StridedCuMatrix{$elty})
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             nA, nB  = checksquare(A, B)
             if nB != nA
                 throw(DimensionMismatch("Dimensions of A ($nA, $nA) and B ($nB, $nB) must match!"))
@@ -589,7 +589,7 @@ for (jname, bname, fname, elty, relty) in ((:sygvj!, :cusolverDnSsygvj_bufferSiz
                         B::StridedCuMatrix{$elty};
                         tol::$relty=eps($relty),
                         max_sweeps::Int=100)
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             nA, nB  = checksquare(A, B)
             if nB != nA
                 throw(DimensionMismatch("Dimensions of A ($nA, $nA) and B ($nB, $nB) must match!"))
@@ -643,7 +643,7 @@ for (jname, bname, fname, elty, relty) in ((:syevjBatched!, :cusolverDnSsyevjBat
                         max_sweeps::Int=100)
 
             # Set up information for the solver arguments
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             n       = checksquare(A)
             lda     = max(1, stride(A, 2))
             batchSize = size(A,3)
@@ -702,7 +702,7 @@ for (fname, elty) in ((:cusolverDnSpotrsBatched, :Float32),
                 throw(DimensionMismatch(""))
             end
             # Set up information for the solver arguments
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             n = checksquare(A[1])
             if size(B[1], 1) != n
                 throw(DimensionMismatch("first dimension of B[i], $(size(B[1],1)), must match second dimension of A, $n"))
@@ -741,7 +741,7 @@ for (fname, elty) in ((:cusolverDnSpotrfBatched, :Float32),
         function potrfBatched!(uplo::Char, A::Vector{<:StridedCuMatrix{$elty}})
 
             # Set up information for the solver arguments
-            VERSION >= v"1.8" && LinearAlgebra.BLAS.chkuplo(uplo)
+            LinearAlgebra.BLAS.chkuplo(uplo)
             n = checksquare(A[1])
             lda = max(1, stride(A[1], 2))
             batchSize = length(A)