Merge pull request #327 from JuliaGPU/tb/1.11

Support for Julia 1.11

Author: maleadt

.buildkite/pipeline.yml

@@ -28,10 +28,10 @@ steps:
               - "1.9"
               - "1.10"
               - "1.11"
-              # - "nightly"
+              - "nightly"
           adjustments:
             - with:
-                julia: "1.11"
+                julia: "nightly"
               soft_fail: true
 
   # Test Storage modes Shared and Managed

lib/mps/linalg.jl

@@ -202,8 +202,21 @@ function LinearAlgebra.lu(A::MtlMatrix{T}; check::Bool = true) where {T<:MtlFloa
     return LinearAlgebra.LU(B, p, status)
 end
 
+function _check_lu_success(info, allowsingular)
+    if VERSION >= v"1.11.0-DEV.1535"
+        if info < 0 # zero pivot error from unpivoted LU
+            LinearAlgebra.checknozeropivot(-info)
+        else
+            allowsingular || LinearAlgebra.checknonsingular(info)
+        end
+    else
+        LinearAlgebra.checknonsingular(info)
+    end
+end
+
 # TODO: dispatch on pivot strategy
-function LinearAlgebra.lu!(A::MtlMatrix{T}; check::Bool = true) where {T<:MtlFloat}
+function LinearAlgebra.lu!(A::MtlMatrix{T};
+                           check::Bool=true, allowsingular::Bool=false) where {T<:MtlFloat}
     M,N = size(A)
     dev = current_device()
     queue = global_queue(dev)
@@ -238,7 +251,7 @@ function LinearAlgebra.lu!(A::MtlMatrix{T}; check::Bool = true) where {T<:MtlFlo
     wait_completed(cmdbuf_lu)
 
     status = convert(LinearAlgebra.BlasInt, Metal.@allowscalar status[])
-    check && checknonsingular(status)
+    check && _check_lu_success(status, allowsingular)
 
     return LinearAlgebra.LU(A, p, status)
 end

src/compiler/compilation.jl

@@ -6,9 +6,6 @@ const MetalCompilerJob = CompilerJob{MetalCompilerTarget, MetalCompilerParams}
 
 GPUCompiler.runtime_module(::MetalCompilerJob) = Metal
 
-const ci_cache = GPUCompiler.CodeCache()
-GPUCompiler.ci_cache(::MetalCompilerJob) = ci_cache
-
 GPUCompiler.method_table(::MetalCompilerJob) = method_table
 
 
@@ -72,6 +69,9 @@ function compile(@nospecialize(job::CompilerJob))
             output = Pipe()
 
             cmd = `$(LLVMDowngrader_jll.llvm_as()) --bitcode-version=5.0 -o -`
+            if LLVM.version() >= v"16"
+                cmd = `$cmd --opaque-pointers=0`
+            end
             proc = run(pipeline(cmd, stdout=output, stderr=stderr, stdin=input); wait=false)
             close(output.in)
 

test/setup.jl

@@ -51,8 +51,15 @@ function runtests(f, name)
 
         # process results
         cpu_rss = Sys.maxrss()
-        passes,fails,error,broken,c_passes,c_fails,c_errors,c_broken =
-            Test.get_test_counts(data[1])
+        if VERSION >= v"1.11.0-DEV.1529"
+            tc = Test.get_test_counts(data[1])
+            passes,fails,error,broken,c_passes,c_fails,c_errors,c_broken =
+                tc.passes, tc.fails, tc.errors, tc.broken, tc.cumulative_passes,
+                tc.cumulative_fails, tc.cumulative_errors, tc.cumulative_broken
+        else
+            passes,fails,errors,broken,c_passes,c_fails,c_errors,c_broken =
+                Test.get_test_counts(data[1])
+        end
         if data[1].anynonpass == false
             data = ((passes+c_passes,broken+c_broken),
                     data[2],