Bump CUTENSOR to v1.7. (#1960)

[skip julia]
[skip cuda]
[skip downstream]
[skip benchmarks]
[skip docs]

Author: maleadt

.buildkite/pipeline.yml

@@ -26,7 +26,8 @@ steps:
               # use the CUDA installation from the CI image
               using CUDA
               CUDA.set_runtime_version!("local")'
-        if: build.message !~ /\[skip tests\]/
+        if: build.message !~ /\[skip tests\]/ &&
+            build.message !~ /\[skip julia\]/
         timeout_in_minutes: 120
         matrix:
           setup:
@@ -63,7 +64,9 @@ steps:
         agents:
           queue: "juliagpu"
           cuda: "*"
-        if: build.message !~ /\[skip tests\]/ && !build.pull_request.draft
+        if: build.message !~ /\[skip tests\]/ &&
+            build.message !~ /\[skip cuda\]/ &&
+            !build.pull_request.draft
         timeout_in_minutes: 120
         matrix:
           setup:
@@ -99,6 +102,9 @@ steps:
             - with:
                 cuda: "12.0"
                 package: "cuDNN"
+            - with:
+                cuda: "12.0"
+                package: "cuTENSOR"
         plugins:
           - JuliaCI/julia#v1:
               version: "1.8"
@@ -110,7 +116,9 @@ steps:
         agents:
           queue: "juliagpu"
           cuda: "*"
-        if: build.message !~ /\[skip tests\]/ && !build.pull_request.draft
+        if: build.message !~ /\[skip tests\]/ &&
+            build.message !~ /\[skip subpackages\]/ &&
+            !build.pull_request.draft
         timeout_in_minutes: 120
         commands: |
           julia -e '
@@ -161,7 +169,9 @@ steps:
         agents:
           queue: "juliagpu"
           cuda: "*"
-        if: build.message !~ /\[skip tests\]/ && !build.pull_request.draft
+        if: build.message !~ /\[skip tests\]/ &&
+            build.message !~ /\[skip downstream\]/ &&
+            !build.pull_request.draft
         timeout_in_minutes: 60
 
   - group: ":eyes: Special"
@@ -212,7 +222,8 @@ steps:
         agents:
           queue: "juliagpu"
           cuda: "*"
-        if: build.message !~ /\[skip docs\]/ && !build.pull_request.draft
+        if: build.message !~ /\[skip docs\]/ &&
+            !build.pull_request.draft
         timeout_in_minutes: 30
 
       # XXX: fails often, and is very slow

lib/cutensor/Project.toml

@@ -1,7 +1,7 @@
 name = "cuTENSOR"
 uuid = "011b41b2-24ef-40a8-b3eb-fa098493e9e1"
 authors = ["Tim Besard <tim.besard@gmail.com>"]
-version = "1.0.4"
+version = "1.1.0"
 
 [deps]
 CEnum = "fa961155-64e5-5f13-b03f-caf6b980ea82"
@@ -10,7 +10,7 @@ CUTENSOR_jll = "35b6c64b-1ee1-5834-92a3-3f624899209a"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 
 [compat]
-CUTENSOR_jll = "~1.6"
+CUTENSOR_jll = "~1.7"
 CUDA = "~4.3"
 CEnum = "0.2, 0.3, 0.4"
 julia = "1.6"

lib/cutensor/src/cuTENSOR.jl

@@ -31,28 +31,26 @@ include("wrappers.jl")
 include("interfaces.jl")
 
 # cache for created, but unused handles
-const idle_handles = HandleCache{CuContext,Base.RefValue{cutensorHandle_t}}()
+const idle_handles = HandleCache{CuContext,Ptr{cutensorHandle_t}}()
 
 function handle()
     cuda = CUDA.active_state()
 
     # every task maintains library state per device
-    LibraryState = @NamedTuple{handle::Base.RefValue{cutensorHandle_t}}
+    LibraryState = @NamedTuple{handle::Ptr{cutensorHandle_t}}
     states = get!(task_local_storage(), :cuTENSOR) do
         Dict{CuContext,LibraryState}()
     end::Dict{CuContext,LibraryState}
 
     # get library state
     @noinline function new_state(cuda)
         new_handle = pop!(idle_handles, cuda.context) do
-            handle = Ref{cutensorHandle_t}()
-            cutensorInit(handle)
-            handle
+            cutensorCreate()
         end
 
         finalizer(current_task()) do task
             push!(idle_handles, cuda.context, new_handle) do
-                # cuTENSOR doesn't need to actively destroy its handle
+                cutensorDestroy(new_handle)
             end
         end
 

lib/cutensor/src/libcutensor.jl

@@ -1,6 +1,6 @@
 using CEnum
 
-# cuTENSOR uses CUDA runtime objects, which are compatible with our driver usage
+# CUTENSOR uses CUDA runtime objects, which are compatible with our driver usage
 const cudaStream_t = CUstream
 
 # outlined functionality to avoid GC frame allocation
@@ -14,7 +14,7 @@ end
 
 function check(f, errs...)
     res = retry_reclaim(in((CUTENSOR_STATUS_ALLOC_FAILED, errs...))) do
-        f()
+        return f()
     end
 
     if res != CUTENSOR_STATUS_SUCCESS
@@ -164,9 +164,14 @@ end
 # typedef void ( * cutensorLoggerCallback_t ) ( int32_t logLevel , const char * functionName , const char * message )
 const cutensorLoggerCallback_t = Ptr{Cvoid}
 
-@checked function cutensorInit(handle)
+@checked function cutensorCreate(handle)
     initialize_context()
-    @ccall libcutensor.cutensorInit(handle::Ptr{cutensorHandle_t})::cutensorStatus_t
+    @ccall libcutensor.cutensorCreate(handle::Ptr{Ptr{cutensorHandle_t}})::cutensorStatus_t
+end
+
+@checked function cutensorDestroy(handle)
+    initialize_context()
+    @ccall libcutensor.cutensorDestroy(handle::Ptr{cutensorHandle_t})::cutensorStatus_t
 end
 
 @checked function cutensorHandleDetachPlanCachelines(handle)
@@ -392,12 +397,12 @@ function cutensorGetErrorString(error)
     @ccall libcutensor.cutensorGetErrorString(error::cutensorStatus_t)::Cstring
 end
 
-# no prototype is found for this function at cutensor.h:745:8, please use with caution
+# no prototype is found for this function at cutensor.h:794:8, please use with caution
 function cutensorGetVersion()
     @ccall libcutensor.cutensorGetVersion()::Csize_t
 end
 
-# no prototype is found for this function at cutensor.h:751:8, please use with caution
+# no prototype is found for this function at cutensor.h:800:8, please use with caution
 function cutensorGetCudartVersion()
     @ccall libcutensor.cutensorGetCudartVersion()::Csize_t
 end
@@ -427,7 +432,7 @@ end
     @ccall libcutensor.cutensorLoggerSetMask(mask::Int32)::cutensorStatus_t
 end
 
-# no prototype is found for this function at cutensor.h:799:18, please use with caution
+# no prototype is found for this function at cutensor.h:848:18, please use with caution
 @checked function cutensorLoggerForceDisable()
     initialize_context()
     @ccall libcutensor.cutensorLoggerForceDisable()::cutensorStatus_t
@@ -460,3 +465,8 @@ end
                                                      typeCompute::cutensorComputeType_t,
                                                      workspaceSize::Ptr{UInt64})::cutensorStatus_t
 end
+
+@checked function cutensorInit(handle)
+    initialize_context()
+    @ccall libcutensor.cutensorInit(handle::Ptr{cutensorHandle_t})::cutensorStatus_t
+end

lib/cutensor/src/wrappers.jl

@@ -16,6 +16,14 @@ function cuda_version()
   VersionNumber(major, minor, patch)
 end
 
+function cutensorCreate()
+  handle_ref = Ref{Ptr{cutensorHandle_t}}()
+  check(CUTENSOR_STATUS_NOT_INITIALIZED) do
+    unsafe_cutensorCreate(handle_ref)
+  end
+  handle_ref[]
+end
+
 
 abstract type CuTensorPlan end
 

lib/cutensornet/Project.toml

@@ -1,18 +1,18 @@
+authors = ["Katharine Hyatt <kshyatt@gmail.com>"]
 name = "cuTensorNet"
 uuid = "448d79b3-4b49-4e06-a5ea-00c62c0dc3db"
-authors = ["Katharine Hyatt <kshyatt@gmail.com>"]
 version = "1.0.4"
 
+[compat]
+CEnum = "0.2, 0.3, 0.4"
+CUDA = "~4.3"
+cuQuantum_jll = "~22.11"
+cuTENSOR = "~1.0, ~1.1"
+julia = "1.6"
+
 [deps]
 CEnum = "fa961155-64e5-5f13-b03f-caf6b980ea82"
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
-cuTENSOR = "011b41b2-24ef-40a8-b3eb-fa098493e9e1"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 cuQuantum_jll = "b75408ef-6fdf-5d74-b65a-7df000ad18e6"
-
-[compat]
-cuQuantum_jll = "~22.11"
-cuTENSOR = "~1.0"
-CUDA = "~4.3"
-CEnum = "0.2, 0.3, 0.4"
-julia = "1.6"
+cuTENSOR = "011b41b2-24ef-40a8-b3eb-fa098493e9e1"

res/wrap/Manifest.toml

@@ -46,10 +46,10 @@ uuid = "62b44479-cb7b-5706-934f-f13b2eb2e645"
 version = "8.9.2+0"
 
 [[CUTENSOR_jll]]
-deps = ["Artifacts", "CUDA_Runtime_jll", "CompilerSupportLibraries_jll", "JLLWrappers", "LazyArtifacts", "Libdl", "Pkg", "TOML"]
-git-tree-sha1 = "cd25ea92f25cc47e92e366f2a7a337eae897cd68"
+deps = ["Artifacts", "CUDA_Runtime_jll", "CompilerSupportLibraries_jll", "JLLWrappers", "LazyArtifacts", "Libdl", "TOML"]
+git-tree-sha1 = "d2c1a1de70a8716acaeee873def6608d854212d5"
 uuid = "35b6c64b-1ee1-5834-92a3-3f624899209a"
-version = "1.6.1+1"
+version = "1.7.0+0"
 
 [[Clang]]
 deps = ["CEnum", "Clang_jll", "Downloads", "Pkg", "TOML"]