Merge pull request #730 from JuliaGPU/tb/compute_sanitizer

Switch tests over to compute-sanitizer.

Author: maleadt

.buildkite/pipeline.yml

@@ -27,7 +27,7 @@ steps:
           version: 1.6-nightly
       - JuliaCI/julia-test#v1:
           julia_args: "-g2"
-          test_args: "--memcheck"
+          test_args: "--sanitize --quickfail --jobs=1"
       - JuliaCI/julia-coverage#v1:
           codecov: true
           dirs:
@@ -39,7 +39,9 @@ steps:
       cuda: "11.2"                    # older versions of CUDA have issues
       cap: "recent"                   # test as much as possible
     env:
-      JULIA_CUDA_MEMORY_POOL: 'none'  # CUDA's memory pool requires compute-sanitizer
+      JULIA_CUDA_MEMORY_POOL: 'none'  # compute-sanitizer uses a lot of memory, so we need device_reset!
+      JULIA_CUDA_VERSION: '11.2'
+      JULIA_CUDA_USE_BINARYBUILDER: 'true'
     if: build.message !~ /\[skip tests\]/
     timeout_in_minutes: 120
 

Artifacts.toml

@@ -85,99 +85,102 @@ lazy = true
 
 [[CUDA110]]
 arch = "powerpc64le"
-git-tree-sha1 = "b22672705ca4f00c784a3f9d58619408d4af9de0"
+git-tree-sha1 = "25a70e995c5457a9b3c7dd7ff8a62d14acc2abc5"
 libc = "glibc"
 os = "linux"
 lazy = true
 
     [[CUDA110.download]]
-    sha256 = "e86a67aa8b1b2cd73d78572401efa75f9bb26f6a259f12d0471c64b74fbe204f"
-    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.0.3+3/CUDA.v11.0.3.powerpc64le-linux-gnu.tar.gz"
+    sha256 = "120ee6f20fc3c3c59611cf3c5b1584ed14658bb5d1bf9fd1b25a14182247d262"
+    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.0.3+4/CUDA.v11.0.3.powerpc64le-linux-gnu.tar.gz"
 
 [[CUDA110]]
 arch = "x86_64"
-git-tree-sha1 = "6b1a60793e5e98abdcfc3724cfa22b2a5348dc09"
+git-tree-sha1 = "74e3e04bdbf56ccf276cd8dd896ad07033846fae"
 libc = "glibc"
 os = "linux"
 lazy = true
 
     [[CUDA110.download]]
-    sha256 = "520e690529f67afe6aabdd8d18dc34d18acf5020cb3dc1fd4e904998d9e17aba"
-    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.0.3+3/CUDA.v11.0.3.x86_64-linux-gnu.tar.gz"
+    sha256 = "291e84f0d598ecbcbe438b1d42022583d061ad5f4eece2b1c06d600332b0367e"
+    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.0.3+4/CUDA.v11.0.3.x86_64-linux-gnu.tar.gz"
 
 [[CUDA110]]
 arch = "x86_64"
-git-tree-sha1 = "2d09da4d71a0762750dee0861e28029d38b08d1e"
+git-tree-sha1 = "1ab27f582deafbc99077d540a01141e620620177"
 os = "windows"
 lazy = true
 
     [[CUDA110.download]]
-    sha256 = "d11ca219e9b91725c6677f36b339459d149ffdcfa3f5e51928fb133158caa15a"
-    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.0.3+3/CUDA.v11.0.3.x86_64-w64-mingw32.tar.gz"
+    sha256 = "0ea0100ee7fa6d67c8d63ea44e719d76f6f70ce1ab5f657d7c97f30fae173af5"
+    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.0.3+4/CUDA.v11.0.3.x86_64-w64-mingw32.tar.gz"
+
 
 [[CUDA111]]
 arch = "powerpc64le"
-git-tree-sha1 = "44dba03dc848a148c9d2430354bf7e52e216364c"
+git-tree-sha1 = "8837163c5563af77039b4a04a49b6e2c3f123ab4"
 libc = "glibc"
 os = "linux"
 lazy = true
 
     [[CUDA111.download]]
-    sha256 = "ac85a364080ea8b97e77fb83967046c54099f7c63769577fa39a1311b68add81"
-    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.1.1+2/CUDA.v11.1.1.powerpc64le-linux-gnu.tar.gz"
+    sha256 = "847f43a4f68c2b08c6275c988ff7c7e5414ad477a625ac78f6e4970969fccc48"
+    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.1.1+3/CUDA.v11.1.1.powerpc64le-linux-gnu.tar.gz"
 
 [[CUDA111]]
 arch = "x86_64"
-git-tree-sha1 = "48c41dccb8db0c9aa9483267cb33719207abe4c1"
+git-tree-sha1 = "4670dd02df5210bd53199f14ec9f8cc027d889e0"
 libc = "glibc"
 os = "linux"
 lazy = true
 
     [[CUDA111.download]]
-    sha256 = "b7242ce10b3fb06d886725209d5b19d565c15c7e244eb84b50262f281a04291c"
-    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.1.1+2/CUDA.v11.1.1.x86_64-linux-gnu.tar.gz"
+    sha256 = "84a9574db7bfb0a59dd03ef1a85874d3f33a7686507d89312700f5c519307cba"
+    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.1.1+3/CUDA.v11.1.1.x86_64-linux-gnu.tar.gz"
 
 [[CUDA111]]
 arch = "x86_64"
-git-tree-sha1 = "ad4cf0816c2c327477c512f476649bfde7ada206"
+git-tree-sha1 = "86505c4367204e1769e6341380841f7f589a2f4d"
 os = "windows"
 lazy = true
 
     [[CUDA111.download]]
-    sha256 = "026a92bcb8d7a5ff6f2e6e262ed8d8387164314941f0dc1b3228e383e04a60a0"
-    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.1.1+2/CUDA.v11.1.1.x86_64-w64-mingw32.tar.gz"
+    sha256 = "a56db28c70e9736f9ea024f3afa7fdedf899b7c998808db7d8a368e0a1208ed9"
+    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.1.1+3/CUDA.v11.1.1.x86_64-w64-mingw32.tar.gz"
+
 
 [[CUDA112]]
 arch = "powerpc64le"
-git-tree-sha1 = "3141108f3144f5170dacc12749a61c14101b42b5"
+git-tree-sha1 = "ef3928da3f9b68a5213a93f91da0d27e32c01e50"
 libc = "glibc"
 os = "linux"
 lazy = true
 
     [[CUDA112.download]]
-    sha256 = "d7d6c399c77cabc75f1387869ca8bbef93cb6a745004993b34306e0b23d5bd18"
-    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.2.1+0/CUDA.v11.2.1.powerpc64le-linux-gnu.tar.gz"
+    sha256 = "770235b69868b88e6db4efc30a8659e9708f3b432028e2032ba589cf2c3efaf8"
+    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.2.1+1/CUDA.v11.2.1.powerpc64le-linux-gnu.tar.gz"
 
 [[CUDA112]]
 arch = "x86_64"
-git-tree-sha1 = "43b02b66f55952515d3cc933404d027fb904cd8b"
+git-tree-sha1 = "18f4e83091aec02d8229c2b009a45a5c22b47664"
 libc = "glibc"
 os = "linux"
 lazy = true
 
     [[CUDA112.download]]
-    sha256 = "70089c452bf923c4951048d336ac32ed28ee3672f8667bc7595fdc6190bf1990"
-    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.2.1+0/CUDA.v11.2.1.x86_64-linux-gnu.tar.gz"
+    sha256 = "6da495c82fae19e0aae8691addc72829376547543324358f39e16835cb208e6e"
+    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.2.1+1/CUDA.v11.2.1.x86_64-linux-gnu.tar.gz"
 
 [[CUDA112]]
 arch = "x86_64"
-git-tree-sha1 = "8b7275b36a973e6345a76b2931ddf397228e34ca"
+git-tree-sha1 = "4765905e93e1e93ca8d2eb52a1e8cec5de4627b1"
 os = "windows"
 lazy = true
 
     [[CUDA112.download]]
-    sha256 = "ed69a6b9630fc83e75856486fd157903c6e93e1d70e0fc7e6c67ca0dacea2b15"
-    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.2.1+0/CUDA.v11.2.1.x86_64-w64-mingw32.tar.gz"
+    sha256 = "6dc0ae6aab8b878864bf926fd9446c71f92f689e6115d6dcedc54ac492d30ea3"
+    url = "https://github.com/JuliaBinaryWrappers/CUDA_jll.jl/releases/download/CUDA-v11.2.1+1/CUDA.v11.2.1.x86_64-w64-mingw32.tar.gz"
+
 
 
 # CUDNN

deps/bindeps.jl

@@ -33,7 +33,7 @@ Returns the CUDA release part of the version as returned by [`version`](@ref).
 toolkit_release() = @after_init(VersionNumber(__toolkit_version[].major, __toolkit_version[].minor))
 
 const __nvdisasm = Ref{String}()
-const __memcheck = Ref{Union{Nothing,String}}()
+const __compute_sanitizer = Ref{Union{Nothing,String}}()
 const __libdevice = Ref{String}()
 const __libcudadevrt = Ref{String}()
 const __libcupti = Ref{Union{Nothing,String}}()
@@ -47,10 +47,10 @@ const __libcudnn = Ref{Union{Nothing,String}}(nothing)
 const __libcutensor = Ref{Union{Nothing,String}}(nothing)
 
 nvdisasm() = @after_init(__nvdisasm[])
-function memcheck()
+function compute_sanitizer()
     @after_init begin
-        @assert has_memcheck() "This functionality is unavailabe as CUDA-MEMCHECK is missing."
-        __memcheck[]
+        @assert has_compute_sanitizer() "This functionality is unavailabe as compute-sanitizer is missing."
+        __compute_sanitizer[]
     end
 end
 libdevice() = @after_init(__libdevice[])
@@ -68,8 +68,8 @@ function libnvtx()
     end
 end
 
-export has_memcheck, has_cupti, has_nvtx
-has_memcheck() = @after_init(__memcheck[]) !== nothing
+export has_compute_sanitizer, has_cupti, has_nvtx
+has_compute_sanitizer() = @after_init(__compute_sanitizer[]) !== nothing
 has_cupti() = @after_init(__libcupti[]) !== nothing
 has_nvtx() = @after_init(__libnvtx[]) !== nothing
 
@@ -173,8 +173,7 @@ function use_artifact_cuda()
 
     __nvdisasm[] = artifact_binary(artifact.dir, "nvdisasm")
     @assert isfile(__nvdisasm[])
-    __memcheck[] = artifact_binary(artifact.dir, "cuda-memcheck")
-    @assert isfile(__memcheck[])
+    __compute_sanitizer[] = artifact_binary(artifact.dir, "compute-sanitizer")
 
     __libcupti[] = artifact_cuda_library(artifact.dir, "cupti", artifact.version)
     @assert isfile(__libcupti[])
@@ -221,7 +220,7 @@ function use_local_cuda()
         __nvdisasm[] = path
     end
 
-    __memcheck[] = find_cuda_binary("cuda-memcheck", cuda_dirs)
+    __compute_sanitizer[] = find_cuda_binary("compute-sanitizer", cuda_dirs)
 
     cuda_version = parse_toolkit_version("nvdisasm", __nvdisasm[])
     if cuda_version === nothing

src/pool.jl

@@ -223,6 +223,9 @@ const pools = PerDevice{AbstractPool}(dev->begin
   pool
 end)
 
+# NVIDIA bug #3240770
+@memoize any_stream_ordered() = any(dev->pools[dev].stream_ordered, devices())
+
 
 ## interface
 

src/state.jl

@@ -344,8 +344,7 @@ so it is generally not needed to subscribe to the reset hook specifically.
     this package.
 """
 function device_reset!(dev::CuDevice=device())
-    stream_ordered = any(dev->pools[dev].stream_ordered, devices())
-    if stream_ordered # NVIDIA bug #3240770
+    if any_stream_ordered()
         @error """Due to a bug in CUDA, resetting the device is not possible on CUDA 11.2 when using the stream-ordered memory allocator.
 
                   If you are calling this function to free memory, that may not be required anymore

test/codegen.jl

@@ -140,22 +140,22 @@ end
     valid_kernel() = return
     invalid_kernel() = 1
 
-    @not_if_memcheck @test CUDA.code_sass(devnull, valid_kernel, Tuple{}) == nothing
-    @not_if_memcheck @test_throws CUDA.KernelError CUDA.code_sass(devnull, invalid_kernel, Tuple{})
+    @not_if_sanitize @test CUDA.code_sass(devnull, valid_kernel, Tuple{}) == nothing
+    @not_if_sanitize @test_throws CUDA.KernelError CUDA.code_sass(devnull, invalid_kernel, Tuple{})
 end
 
 @testset "function name mangling" begin
     @eval @noinline $(Symbol("dummy_^"))(x) = x
 
     @eval kernel_341(ptr) = (@inbounds unsafe_store!(ptr, $(Symbol("dummy_^"))(unsafe_load(ptr))); nothing)
 
-    @not_if_memcheck CUDA.code_sass(devnull, kernel_341, Tuple{Ptr{Int}})
+    @not_if_sanitize CUDA.code_sass(devnull, kernel_341, Tuple{Ptr{Int}})
 end
 
 @testset "device runtime" begin
     kernel() = (CUDA.cudaGetLastError(); return)
 
-    @not_if_memcheck CUDA.code_sass(devnull, kernel, Tuple{})
+    @not_if_sanitize CUDA.code_sass(devnull, kernel, Tuple{})
 end
 
 end

test/cudadrv/module.jl

@@ -24,7 +24,7 @@ let
     @test md != md2
 end
 
-@not_if_memcheck @test_throws_cuerror CUDA.ERROR_INVALID_IMAGE CuModule("foobar")
+@not_if_sanitize @test_throws_cuerror CUDA.ERROR_INVALID_IMAGE CuModule("foobar")
 
 @testset "globals" begin
     md = CuModuleFile(joinpath(@__DIR__, "ptx/global.ptx"))
@@ -54,11 +54,11 @@ end
     # TODO: test with valid object code
     # NOTE: apparently, on Windows cuLinkAddData! _does_ accept object data containing \0
     if !Sys.iswindows()
-        @not_if_memcheck @test_throws_cuerror CUDA.ERROR_UNKNOWN add_data!(link, "vadd_parent", UInt8[0])
+        @not_if_sanitize @test_throws_cuerror CUDA.ERROR_UNKNOWN add_data!(link, "vadd_parent", UInt8[0])
     end
 end
 
-@not_if_memcheck @testset "error log" begin
+@not_if_sanitize @testset "error log" begin
     @test_throws_message contains("ptxas fatal") CuError CuModule(".version 3.1")
 
     link = CuLink()

test/cudadrv/pool.jl

@@ -1,4 +1,4 @@
-@not_if_memcheck let
+@not_if_sanitize let
     dev = device()
 
     pool = memory_pool(dev)

test/cutensor/contractions.jl

@@ -2,6 +2,9 @@ using CUDA.CUTENSOR
 using CUDA
 using LinearAlgebra
 
+# these tests perform a lot of harmless-but-invalid API calls, poluting sanitizer logs
+@not_if_sanitize begin
+
 eltypes = ( (Float32, Float32, Float32, Float32),
             (Float32, Float32, Float32, Float16),
             (ComplexF32, ComplexF32, ComplexF32, ComplexF32),
@@ -196,3 +199,5 @@ can_pin = !Sys.iswindows()
         end
     end
 end
+
+end

test/examples.jl

@@ -1,5 +1,9 @@
+# NVIDIA bug 3263616: compute-sanitizer crashes when generating host backtraces,
+#                     but --show-backtrace=no does not survive execve.
+@not_if_sanitize begin
+
 # these tests spawn subprocesses, so reset the current context to conserve memory
-CUDA.release() == v"11.2" || CUDA.device_reset!()
+CUDA.any_stream_ordered() || CUDA.device_reset!()
 
 function find_sources(path::String, sources=String[])
     if isdir(path)
@@ -28,3 +32,5 @@ cd(examples_dir) do
         @test success(pipeline(`$cmd $example`, stderr=stderr))
     end
 end
+
+end