Merge branch 'master' into qr_views

Author: evelyne-ringoot

.buildkite/pipeline.yml

@@ -145,8 +145,8 @@ steps:
     steps:
       - label: "{{matrix}} on CUDA 10"
         matrix:
-            - "CUDNN"
-            - "CUTENSOR"
+            - "cuDNN"
+            - "cuTENSOR"
         plugins:
           - JuliaCI/julia#v1:
               version: 1.6
@@ -178,10 +178,10 @@ steps:
             Pkg.test()'
       - label: "{{matrix}} on CUDA 11"
         matrix:
-            - "CUDNN"
-            - "CUTENSOR"
-            - "CUSTATEVEC"
-            - "CUTENSORNET"
+            - "cuDNN"
+            - "cuTENSOR"
+            - "cuStateVec"
+            - "cuTensorNet"
         plugins:
           - JuliaCI/julia#v1:
               version: 1.6
@@ -203,8 +203,8 @@ steps:
 
             println("--- :julia: Instantiating project")
             Pkg.activate(joinpath(pwd(), "lib", lowercase("{{matrix}}")))
-            if "{{matrix}}" == "CUTENSORNET"
-              # HACK: CUTENSORNET depends on a development version of CUTENSOR
+            if "{{matrix}}" == "cuTensorNet"
+              # HACK: cuTensorNet depends on a development version of cuTENSOR
               Pkg.develop(path=joinpath(pwd(), "lib", "cutensor"))
             end
             Pkg.develop(path=pwd())
@@ -279,26 +279,26 @@ steps:
         if: build.message !~ /\[skip tests\]/ && !build.pull_request.draft
         timeout_in_minutes: 60
 
-      - label: "Julia nightly"
-        plugins:
-          - JuliaCI/julia#v1:
-              version: nightly
-          - JuliaCI/julia-test#v1: ~
-          - JuliaCI/julia-coverage#v1:
-              codecov: true
-              dirs:
-                - src
-                - lib
-                - examples
-        agents:
-          queue: "juliagpu"
-          cuda: "*"
-        commands: |
-          julia --project -e 'using CUDA; CUDA.set_runtime_version!("local")'
-        if: build.message !~ /\[skip tests\]/ && !build.pull_request.draft
-        timeout_in_minutes: 120
-        soft_fail:
-          - exit_status: 1
+      # - label: "Julia nightly"
+      #   plugins:
+      #     - JuliaCI/julia#v1:
+      #         version: nightly
+      #     - JuliaCI/julia-test#v1: ~
+      #     - JuliaCI/julia-coverage#v1:
+      #         codecov: true
+      #         dirs:
+      #           - src
+      #           - lib
+      #           - examples
+      #   agents:
+      #     queue: "juliagpu"
+      #     cuda: "*"
+      #   commands: |
+      #     julia --project -e 'using CUDA; CUDA.set_runtime_version!("local")'
+      #   if: build.message !~ /\[skip tests\]/ && !build.pull_request.draft
+      #   timeout_in_minutes: 120
+      #   soft_fail:
+      #     - exit_status: 1
 
       - label: "Documentation"
         plugins:

Project.toml

@@ -1,6 +1,6 @@
 name = "CUDA"
 uuid = "052768ef-5323-5732-b1bb-66c8b64840ba"
-version = "4.0.0"
+version = "4.0.1"
 
 [deps]
 AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"

docs/src/installation/overview.md

@@ -129,7 +129,7 @@ This generates the following `LocalPreferences.toml` file in your active environ
 version = "11.8"
 ```
 
-This preference is compatible with other CUDA JLLs, e.g., if you load CUDNN_jll it will
+This preference is compatible with other CUDA JLLs, e.g., if you load `CUDNN_jll` it will
 only select artifacts that are compatible with the configured CUDA runtime.
 
 ### Using a local CUDA
@@ -175,7 +175,7 @@ binaries aren't on a globally-discoverable path. For more information, run Julia
 `JULIA_DEBUG` environment variable set to `CUDA_Runtime_Discovery`.
 
 Note that setting the version to `"local"` disables use of *any* CUDA-related JLL, not just
-of CUDA_Runtime_jll. This is out of necessity: JLLs are baked in the precompilation image at
+of `CUDA_Runtime_jll`. This is out of necessity: JLLs are baked in the precompilation image at
 compile time, while local toolkit discovery happens at run time; this inconsistency makes it
 impossible to select a compatible artifact for the JLLs. If you care about other JLLs, use
 CUDA from artifacts.

lib/cublas/CUBLAS.jl

@@ -2,12 +2,13 @@ module CUBLAS
 
 using ..APIUtils
 
-using ..CUDA_Runtime
-
 using ..CUDA
 using ..CUDA: CUstream, cuComplex, cuDoubleComplex, libraryPropertyType, cudaDataType, i32
 using ..CUDA: unsafe_free!, @retry_reclaim, isdebug, @sync, initialize_context
 
+using ..CUDA: CUDA_Runtime
+using ..CUDA_Runtime
+
 using GPUArrays
 
 using LinearAlgebra
@@ -235,6 +236,14 @@ function _log_message(blob)
 end
 
 function __init__()
+    precompiling = ccall(:jl_generating_output, Cint, ()) != 0
+    precompiling && return
+
+    if !CUDA_Runtime.is_available()
+        #@error "cuBLAS is not available"
+        return
+    end
+
     # register a log callback
     log_cond[] = Base.AsyncCondition() do async_cond
         blob = ""

lib/cudadrv/state.jl

@@ -213,9 +213,19 @@ function context(dev::CuDevice)
     end
 
     if capability(dev) < v"3.5"
-        @warn("""Your $(name(dev)) GPU does not meet the minimal required compute capability ($(capability(dev)) < 3.5).
-                 Some functionality might be unavailable.""",
-              maxlog=1, _id=devidx)
+        @error("""Your $(name(dev)) GPU (compute capability $(capability(dev).major).$(capability(dev).minor)) is not supported by CUDA.jl.
+                  Please use a device with at least capability 3.5.""",
+               maxlog=1, _id=devidx)
+    elseif runtime_version() >= v"12" && capability(dev) <= v"3.7"
+        @error("""Your $(name(dev)) GPU (compute capability $(capability(dev).major).$(capability(dev).minor)) is not supported on CUDA 12+.
+                  Please use CUDA 11.8 or earlier (by calling `CUDA.set_runtime_version!`), or switch to a different device.""",
+               maxlog=1, _id=devidx)
+    elseif runtime_version() >= v"11" && capability(dev) <= v"3.7"
+        # XXX: the 10.2 release notes mention that even sm_50 is deprecated,
+        #      but that isn't repeated in more recent release notes...
+        @warn("""Your $(name(dev)) GPU (compute capability $(capability(dev).major).$(capability(dev).minor)) is deprecated on CUDA 11+.
+                  Some functionality may be broken; It's recommended to switch to a different device.""",
+               maxlog=1, _id=devidx)
     end
 
     # configure the primary context

lib/cudnn/LICENSE.md

@@ -0,0 +1,28 @@
+The CUDA.jl package is licensed under the MIT "Expat" License:
+
+> Copyright (c) 2019-present: Julia Computing and other contributors
+> 
+> Copyright (c) 2014-2018: Tim Besard
+> 
+> Copyright (c) 2013: Dahua Lin
+>
+> All Rights Reserved.
+>
+> Permission is hereby granted, free of charge, to any person obtaining a copy
+> of this software and associated documentation files (the "Software"), to deal
+> in the Software without restriction, including without limitation the rights
+> to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+> copies of the Software, and to permit persons to whom the Software is
+> furnished to do so, subject to the following conditions:
+>
+> The above copyright notice and this permission notice shall be included in all
+> copies or substantial portions of the Software.
+>
+> THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+> IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+> FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+> AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+> LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+> OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+> SOFTWARE.
+>

lib/cudnn/Project.toml

@@ -1,13 +1,15 @@
-name = "CUDNN"
+name = "cuDNN"
 uuid = "02a925ec-e4fe-4b08-9a7e-0d78e3d38ccd"
 authors = ["Tim Besard <tim.besard@gmail.com>"]
-version = "1.0.0"
+version = "1.0.1"
 
 [deps]
 CEnum = "fa961155-64e5-5f13-b03f-caf6b980ea82"
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 CUDNN_jll = "62b44479-cb7b-5706-934f-f13b2eb2e645"
-Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 
 [compat]
 CUDNN_jll = "~8.6"
+CUDA = "~4.0"
+CEnum = "0.2, 0.3, 0.4"
+julia = "1.6"

lib/cudnn/src/convolution.jl

@@ -40,7 +40,7 @@ Keyword arguments describing the convolution when `d` is not given:
 * `stride = 1`: how far to shift the convolution window at each step
 * `dilation = 1`: dilation factor
 * `group = 1`: number of groups to be used
-* `mathType = CUDNN.math_mode()`: whether or not the use of tensor op is permitted
+* `mathType = cuDNN.math_mode()`: whether or not the use of tensor op is permitted
 * `reorderType = CUDNN_DEFAULT_REORDER`: convolution reorder type
 """
 cudnnConvolutionForward, cudnnConvolutionForward!

lib/cudnn/src/CUDNN.jl renamed to lib/cudnn/src/cuDNN.jl

@@ -1,11 +1,11 @@
 """
-    CUDNN
+    cuDNN
 
 High level interface to cuDNN functions. See
 [README.md](https://github.com/JuliaGPU/CUDA.jl/blob/master/lib/cudnn/README.md) for a
 design overview.
 """
-module CUDNN
+module cuDNN
 
 using CUDA
 using CUDA.APIUtils
@@ -20,8 +20,11 @@ using CUDNN_jll
 export has_cudnn
 
 function has_cudnn(show_reason::Bool=false)
-    if !isdefined(CUDNN_jll, :libcudnn)
-        show_reason && error("CUDNN library not found")
+    precompiling = ccall(:jl_generating_output, Cint, ()) != 0
+    precompiling && return
+
+    if !CUDNN_jll.is_available()
+        show_reason && error("cuDNN library not found")
         return false
     end
     return true
@@ -72,7 +75,7 @@ function handle()
 
     # every task maintains library state per device
     LibraryState = @NamedTuple{handle::cudnnHandle_t, stream::CuStream}
-    states = get!(task_local_storage(), :CUDNN) do
+    states = get!(task_local_storage(), :cuDNN) do
         Dict{CuContext,LibraryState}()
     end::Dict{CuContext,LibraryState}
 
@@ -155,13 +158,16 @@ function _log_message(sev, dbg, str)
 end
 
 function __init__()
-    if version() < v"8.0"
-        @warn "This version of CUDA.jl only supports CUDNN 8.0 or higher"
+    precompiling = ccall(:jl_generating_output, Cint, ()) != 0
+    precompiling && return
+
+    if !CUDNN_jll.is_available()
+        #@error "cuDNN is not available for your platform ($(Base.BinaryPlatforms.triplet(CUDNN_jll.host_platform)))"
+        return
     end
 
     # register a log callback
-    if (isdebug(:init, CUDNN) || Base.JLOptions().debug_level >= 2) &&
-       version() >= v"8.2"  # NVIDIA bug #3256123
+    if isdebug(:init, cuDNN) || Base.JLOptions().debug_level >= 2
         log_cond[] = Base.AsyncCondition() do async_cond
             message = Base.@lock log_lock popfirst!(log_messages)
             _log_message(message...)

lib/cudnn/src/dropout.jl

@@ -13,7 +13,7 @@ the same values deterministically for debugging. Note that this slows down the o
 about 40x.
 
 The global constant `cudnnDropoutState::Dict` holds the random number generator state for
-each CUDNN handle.
+each cuDNN handle.
 """
 cudnnDropoutForward, cudnnDropoutForward!, cudnnDropoutSeed, cudnnDropoutState