Add 1.11 support (#635)

.buildkite/pipeline.yml

@@ -42,6 +42,47 @@ steps:
       JULIA_AMDGPU_HIP_MUST_LOAD: "1"
       JULIA_AMDGPU_DISABLE_ARTIFACTS: "1"
 
+  - label: "Julia 1.11 typed pointers"
+    plugins:
+      - JuliaCI/julia#v1:
+          version: "1.11"
+      - JuliaCI/julia-test#v1:
+      - JuliaCI/julia-coverage#v1:
+          codecov: true
+    agents:
+      queue: "juliagpu"
+      rocm: "*"
+      rocmgpu: "*"
+    if: build.message !~ /\[skip tests\]/
+    command: "julia --project -e 'using Pkg; Pkg.update()'"
+    timeout_in_minutes: 180
+    env:
+      JULIA_NUM_THREADS: 4
+      JULIA_AMDGPU_CORE_MUST_LOAD: "1"
+      JULIA_AMDGPU_HIP_MUST_LOAD: "1"
+      JULIA_AMDGPU_DISABLE_ARTIFACTS: "1"
+
+  - label: "Julia 1.11 opaque pointers"
+    plugins:
+      - JuliaCI/julia#v1:
+          version: "1.11"
+      - JuliaCI/julia-test#v1:
+      - JuliaCI/julia-coverage#v1:
+          codecov: true
+    agents:
+      queue: "juliagpu"
+      rocm: "*"
+      rocmgpu: "*"
+    if: build.message !~ /\[skip tests\]/
+    command: "julia --project -e 'using Pkg; Pkg.update()'"
+    timeout_in_minutes: 180
+    env:
+      JULIA_LLVM_ARGS: "-opaque-pointers"
+      JULIA_NUM_THREADS: 4
+      JULIA_AMDGPU_CORE_MUST_LOAD: "1"
+      JULIA_AMDGPU_HIP_MUST_LOAD: "1"
+      JULIA_AMDGPU_DISABLE_ARTIFACTS: "1"
+
   - label: "GPU-less environment"
     plugins:
       - JuliaCI/julia#v1:

.gitignore

@@ -8,5 +8,5 @@
 
 docs/build
 docs/Manifest.toml
-Manifest.toml
+Manifest*.toml
 LocalPreferences.toml

src/AMDGPU.jl

@@ -74,7 +74,6 @@ import .Runtime: Mem, ROCDim, ROCDim3
 
 include("memory.jl")
 
-const ci_cache = GPUCompiler.CodeCache()
 Base.Experimental.@MethodTable(method_table)
 
 # Device sources must load _before_ the compiler infrastructure,

src/compiler/codegen.jl

@@ -23,8 +23,6 @@ end
 
 GPUCompiler.runtime_module(@nospecialize(::HIPCompilerJob)) = AMDGPU
 
-GPUCompiler.ci_cache(@nospecialize(::HIPCompilerJob)) = AMDGPU.ci_cache
-
 GPUCompiler.method_table(@nospecialize(::HIPCompilerJob)) = AMDGPU.method_table
 
 GPUCompiler.kernel_state_type(@nospecialize(::HIPCompilerJob)) = AMDGPU.KernelState

src/compiler/zeroinit_lds.jl

@@ -22,9 +22,11 @@ function zeroinit_lds!(mod::LLVM.Module, entry::LLVM.Function)
     to_init = []
     for gbl in LLVM.globals(mod)
         if startswith(LLVM.name(gbl), "__zeroinit")
-            as = LLVM.addrspace(value_type(gbl))
+            vt = value_type(gbl)
+            as = LLVM.addrspace(vt)
             if as == AMDGPU.Device.AS.Local
-                push!(to_init, gbl)
+                sz = llvmsize(global_value_type(gbl))
+                push!(to_init, (gbl, sz))
             end
         end
     end
@@ -37,13 +39,10 @@ function zeroinit_lds!(mod::LLVM.Module, entry::LLVM.Function)
         position!(builder, instruction)
 
         # Use memset to clear all values to 0.
-        for gbl in to_init
-            sz = llvmsize(eltype(value_type(gbl)))
+        for (gbl, sz) in to_init
             sz == 0 && continue
-
             LLVM.memset!(builder, gbl,
-                ConstantInt(UInt8(0)), ConstantInt(sz),
-                LLVM.alignment(gbl))
+                ConstantInt(UInt8(0)), ConstantInt(sz), LLVM.alignment(gbl))
         end
 
         # Synchronize the workgroup to prevent races.

src/device/gcn/assertion.jl

@@ -29,10 +29,11 @@ macro rocassert(ex, msgs...)
         msg = :(Main.Base.string($(Expr(:quote,msg))))
     end
 
-    return :($(esc(ex)) ? $(nothing)
-                        : rocassert_fail($(Val(Symbol(msg))),
-                                         $(Val(__source__.file)),
-                                         $(Val(__source__.line))))
+    return :($(esc(ex)) ? $(nothing) :
+        rocassert_fail(
+            $(Val(Symbol(msg))),
+            $(Val(__source__.file)),
+            $(Val(__source__.line))))
 end
 
 assert_counter = 0

src/device/gcn/math.jl

@@ -22,8 +22,11 @@ for jltype in (Float64, Float32, Float16)
     type_suffix = fntypes[jltype]
 
     for (mod, intrinsic) in DEFINED_UNARY_INTRNISICS
-        # sin(Float16) is broken, we override it manually at the very bottom.
+        # sin(::Float16) is broken, we override it manually at the very bottom.
         jltype == Float16 && intrinsic == :sin && continue
+        # log(::Float16) is broken, we override it manually at the very bottom.
+        # https://github.com/ROCm/llvm-project/blob/592734c97a3ddcb7ca4009ac94550595a52450ce/amd/device-libs/ocml/src/logH.cl#L15
+        jltype == Float16 && intrinsic == :log && continue
 
         fname = "extern __ocml_$(intrinsic)_$(type_suffix)"
         if isnothing(mod)
@@ -132,4 +135,5 @@ end
 end
 
 # sin(Float16) is broken, so cast to Float32, see #177.
-@device_override Base.sin(x::Float16) = sin(Float32(x))
+@device_override Base.sin(x::Float16) = Float16(sin(Float32(x)))
+@device_override Base.log(x::Float16) = Float16(log(Float32(x)))

src/device/gcn/memory_static.jl

@@ -42,8 +42,7 @@
             entry = BasicBlock(llvm_f, "entry")
             position!(builder, entry)
 
-            ptr_with_as = gep!(builder, gv_typ, gv,
-                [ConstantInt(0), ConstantInt(0)])
+            ptr_with_as = gep!(builder, gv_typ, gv, [ConstantInt(0), ConstantInt(0)])
             ptr = bitcast!(builder, ptr_with_as, T_ptr_i8)
             ret!(builder, ptr)
         end

src/device/random.jl

@@ -63,27 +63,18 @@ struct Philox2x32{R} <: RandomNumbers.AbstractRNG{UInt64}
     @inline function Philox2x32{R}() where R
         rng = new{R}()
         if rng.key == 0
-            # initialize the key. this happens when first accessing the (0-initialized)
-            # shared memory key from each block. if we ever want to make the device seed
-            # controlable from the host, this would be the place to read a global seed.
-            #
-            # note however that it is undefined how shared memory persists across e.g.
-            # launches, so we may not be able to rely on the zero initalization then.
-            rng.key = Random.make_seed()
+            # initialize the key. this happens when first accessing
+            # the (0-initialized) shared memory key from each block.
+            @static if VERSION >= v"1.11-"
+                Random.seed!(rng, nothing)
+            else
+                rng.key = Random.make_seed()
+            end
         end
         return rng
     end
 end
 
-if VERSION >= v"1.11-"
-    # `Random.seed!(::AbstractRNG)` now passes a `nothing` seed value
-    Random.seed!(rng::Philox2x32, seed::Nothing) =
-        Random.seed!(rng, Base.unsafe_trunc(UInt32, readcyclecounter()))
-else
-    # ... where it used to call `Random_make_seed()`
-    @device_override Random.make_seed() = Base.unsafe_trunc(UInt32, readcyclecounter())
-end
-
 # default to 7 rounds; enough to pass SmallCrush
 @inline Philox2x32() = Philox2x32{7}()
 
@@ -135,6 +126,7 @@ function Random.seed!(rng::Philox2x32, seed::Integer, counter::Integer=0)
     rng.ctr1 = counter
     return
 end
+
 # seeding the implicit default RNG
 if VERSION >= v"1.11-"
     @device_override Random.seed!(seed) =
@@ -144,6 +136,15 @@ else
         Random.seed!(Random.default_rng(), seed)
 end
 
+if VERSION >= v"1.11-"
+    # `Random.seed!(::AbstractRNG)` now passes a `nothing` seed value
+    Random.seed!(rng::Philox2x32, seed::Nothing) =
+        Random.seed!(rng, Base.unsafe_trunc(UInt32, readcyclecounter()))
+else
+    # ... where it used to call `Random_make_seed()`
+    @device_override Random.make_seed() = Base.unsafe_trunc(UInt32, readcyclecounter())
+end
+
 """
     Random.rand(rng::Philox2x32, UInt32)
 

src/discovery/discovery.jl

@@ -142,7 +142,10 @@ function __init__()
         else
             # Detect HIP version, which will influence what device libraries to use.
             hip_version = Base.thisminor(_hip_runtime_version())
-            hip_version > v"5.4" ? true : use_artifacts()
+            # Check if opaque pointers are enabled and turn off artifacts.
+            llvm_args = get(ENV, "JULIA_LLVM_ARGS", "")
+            enabled_opaque_pointers = occursin("-opaque-pointers", llvm_args)
+            (hip_version > v"5.4" && !enabled_opaque_pointers) ? true : use_artifacts()
         end
         # If ROCm 5.5+ - use artifact device libraries.
         global libdevice_libs = get_device_libs(from_artifact; rocm_path)

src/kernels/indexing.jl

@@ -1,6 +1,9 @@
 Base.to_index(::ROCArray, I::AbstractArray{Bool}) = findall(I)
 
-# TODO Julia 1.11 specifics
+if VERSION >= v"1.11.0-DEV.1157"
+    Base.to_indices(x::ROCArray, I::Tuple{AbstractArray{Bool}}) =
+        (Base.to_index(x, I[1]),)
+end
 
 function Base.findall(bools::AnyROCArray{Bool})
     I = keytype(bools)