Support Julia 1.13

.buildkite/pipeline.yml

@@ -33,6 +33,7 @@ steps:
               - "1.10"
               - "1.11"
               - "1.12"
+              - "1.13"
               - "nightly"
           adjustments:
             - with:

lib/nvml/NVML.jl

@@ -12,21 +12,17 @@ import Libdl
 
 export has_nvml
 
-function libnvml()
-    @memoize begin
-        if Sys.iswindows()
-            # the NVSMI dir isn't added to PATH by the installer
-            nvsmi = joinpath(ENV["ProgramFiles"], "NVIDIA Corporation", "NVSMI")
-            if isdir(nvsmi)
-                joinpath(nvsmi, "nvml.dll")
-            else
-                # let's just hope for the best
-                "nvml"
-            end
-        else
-            "libnvidia-ml.so.1"
-        end
-    end::String
+const libnvml::String = if Sys.iswindows()
+    # the NVSMI dir isn't added to PATH by the installer
+    nvsmi = joinpath(ENV["ProgramFiles"], "NVIDIA Corporation", "NVSMI")
+    if isdir(nvsmi)
+        joinpath(nvsmi, "nvml.dll")
+    else
+        # let's just hope for the best
+        "nvml"
+    end
+else
+    "libnvidia-ml.so.1"
 end
 
 function has_nvml()
@@ -37,7 +33,7 @@ function has_nvml()
             return false
         end
 
-        if Libdl.dlopen(libnvml(); throw_error=false) === nothing
+        if Libdl.dlopen(libnvml; throw_error=false) === nothing
             return false
         end
 

res/wrap/nvml.toml

@@ -1,5 +1,5 @@
 [general]
-library_name = "libnvml()"
+library_name = "libnvml"
 output_file_path = "../../lib/nvml/libnvml.jl"
 prologue_file_path = "./libnvml_prologue.jl"
 

src/device/intrinsics/indexing.jl

@@ -63,54 +63,54 @@ end
 
 @device_functions begin
 
-"""
-    gridDim()::NamedTuple
+@doc """
+    threadIdx()::NamedTuple
 
-Returns the dimensions of the grid.
-"""
-@inline gridDim() =   (x=gridDim_x(),   y=gridDim_y(),   z=gridDim_z())
+Returns the thread index within the block.
+""" threadIdx
+@inline threadIdx() = (x=threadIdx_x(), y=threadIdx_y(), z=threadIdx_z())
 
-"""
-    blockIdx()::NamedTuple
+@doc """
+    blockDim()::NamedTuple
 
-Returns the block index within the grid.
-"""
-@inline blockIdx() =  (x=blockIdx_x(),  y=blockIdx_y(),  z=blockIdx_z())
+Returns the dimensions (in threads) of the block.
+""" blockDim
+@inline blockDim() = (x=blockDim_x(), y=blockDim_y(), z=blockDim_z())
 
-"""
-    blockDim()::NamedTuple
+@doc """
+    blockIdx()::NamedTuple
 
-Returns the dimensions of the block.
-"""
-@inline blockDim() =  (x=blockDim_x(),  y=blockDim_y(),  z=blockDim_z())
+Returns the block index within the grid.
+""" blockIdx
+@inline blockIdx() = (x=blockIdx_x(), y=blockIdx_y(), z=blockIdx_z())
 
-"""
-    threadIdx()::NamedTuple
+@doc """
+    gridDim()::NamedTuple
 
-Returns the thread index within the block.
-"""
-@inline threadIdx() = (x=threadIdx_x(), y=threadIdx_y(), z=threadIdx_z())
+Returns the dimensions (in blocks) of the grid.
+""" gridDim
+@inline gridDim() = (x=gridDim_x(), y=gridDim_y(), z=gridDim_z())
 
-"""
+@doc """
     warpsize()::Int32
 
 Returns the warp size (in threads).
-"""
+""" warpsize
 @inline warpsize() = ccall("llvm.nvvm.read.ptx.sreg.warpsize", llvmcall, Int32, ())
 
-"""
+@doc """
     laneid()::Int32
 
 Returns the thread's lane within the warp.
-"""
+""" laneid
 @inline laneid() = ccall("llvm.nvvm.read.ptx.sreg.laneid", llvmcall, Int32, ()) + 1i32
 
-"""
+@doc """
     lanemask(pred)::UInt32
 
 Returns a 32-bit mask indicating which threads in a warp satisfy the given predicate.
 Supported predicates are `==`, `<`, `<=`, `>=`, and `>`.
-"""
+""" lanemask
 @inline function lanemask(pred::F) where F
     if pred === Base.:(==)
         ccall("llvm.nvvm.read.ptx.sreg.lanemask.eq", llvmcall, UInt32, ())
@@ -127,12 +127,12 @@ Supported predicates are `==`, `<`, `<=`, `>=`, and `>`.
     end
 end
 
-"""
+@doc """
     active_mask()
 
 Returns a 32-bit mask indicating which threads in a warp are active with the current
 executing thread.
-"""
+""" active_mask
 @inline active_mask() = @asmcall("activemask.b32 \$0;", "=r", false, UInt32, Tuple{})
 
 end

src/device/utils.jl

@@ -52,8 +52,12 @@ macro device_functions(ex)
                 # descend in blocks
                 push!(out.args, rewrite(arg))
             elseif Meta.isexpr(arg, [:function, :(=)])
-                # rewrite function definitions
-                push!(out.args, :(@device_function $arg))
+                # capture temp variable for Julia 1.13 and rewrite function definitions
+                if Meta.isexpr(arg, :(=)) && isa(arg.args[1], Symbol) && Meta.isexpr(arg.args[2], [:function, :(=)])
+                    push!(out.args, Expr(:(=), arg.args[1], :(@device_function $(arg.args[2]))))
+                else
+                    push!(out.args, :(@device_function $arg))
+                end
             else
                 # preserve all the rest
                 push!(out.args, arg)

test/base/texture.jl

@@ -1,3 +1,6 @@
+# Broken on LLVM 20, see JuliaGPU/CUDA.jl#3037
+if !(v"20-" <= Base.libllvm_version < v"21-")
+
 using Interpolations
 
 @inline function calcpoint(blockIdx, blockDim, threadIdx, size)
@@ -241,3 +244,5 @@ end
         end
     end
 end
+
+end

test/core/device/ldg.jl

@@ -1,6 +1,12 @@
 @testset "ldg" begin
-    ir = sprint(io->CUDA.code_llvm(io, CUDA.pointerref_ldg, Tuple{Core.LLVMPtr{Int,AS.Global},Int,Val{1}}))
-    @test occursin("@llvm.nvvm.ldg", ir)
+    ir = sprint(io->CUDA.code_llvm(io, CUDA.pointerref_ldg, Tuple{Core.LLVMPtr{Int,AS.Global},Int,Val{1}}; raw=true))
+    if Base.libllvm_version >= v"20"
+        # `@llvm.nvvm.ldg` was removed in LLVM 20; the auto-upgrade
+        # replaces it with a load bearing `!invariant.load` metadata
+        @test occursin("!invariant.load", ir)
+    else
+        @test occursin("@llvm.nvvm.ldg", ir)
+    end
 end