Fails to create array on Rusticl driver

[VarLad]

Tried with AMD Vega8 GPU as well as RTX3050M
The command to run julia:

RUSTICL_ENABLE=zink julia

and to reproduce:

julia> using OpenCL

julia> OpenCL.versioninfo()
OpenCL.jl version 0.10.0

Toolchain:
 - Julia v1.11.0-rc4
 - OpenCL_jll v2024.5.8+1

Available platforms: 2
 - Clover
   OpenCL 1.1, Mesa 24.2.4
   · AMD Radeon Graphics (radeonsi, renoir, LLVM 19.1.0, DRM 3.59, 6.11.2-300.fc41.x86_64) (fp64)
 - rusticl
   OpenCL 3.0
   · zink Vulkan 1.3(NVIDIA GeForce RTX 3050 Laptop GPU (NVK GA107) (MESA_NVK)) (il)
   · zink Vulkan 1.3(AMD Radeon Graphics (RADV RENOIR) (MESA_RADV)) (il)

julia> OpenCL.cl.device!(OpenCL.cl.devices(OpenCL.cl.platform!("rusticl"))[1])
┌ Warning: No OpenCL driver JLLs were detected at the time of the first call into OpenCL.jl.
│ Only system drivers will be available.
└ @ OpenCL.cl ~/.julia/packages/OpenCL/VxZWM/lib/cl/api.jl:80
OpenCL.Device(zink Vulkan 1.3(NVIDIA GeForce RTX 3050 Laptop GPU (NVK GA107) (MESA_NVK)) on rusticl @0x00000000323d75b8)

julia> CLArray([1,2])
ERROR: CLError(code=-59, CL_INVALID_OPERATION)
Stacktrace:
  [1] throw_api_error(res::Int32)
    @ OpenCL.cl ~/.julia/packages/OpenCL/VxZWM/lib/cl/libopencl.jl:3
  [2] check
    @ ~/.julia/packages/OpenCL/VxZWM/lib/cl/libopencl.jl:14 [inlined]
  [3] clEnqueueSVMMemcpy
    @ ~/.julia/packages/OpenCL/VxZWM/lib/cl/api.jl:34 [inlined]
  [4] enqueue_svm_memcpy(dst::Ptr{Int64}, src::Ptr{Int64}, nbytes::Int64; blocking::Bool, wait_for::Vector{OpenCL.cl.Event})
    @ OpenCL.cl ~/.julia/packages/OpenCL/VxZWM/lib/cl/svm.jl:63
  [5] enqueue_svm_memcpy
    @ ~/.julia/packages/OpenCL/VxZWM/lib/cl/svm.jl:57 [inlined]
  [6] unsafe_copyto!(dst::CLArray{Int64, 1}, dst_off::Int64, src::Vector{Int64}, src_off::Int64, N::Int64; blocking::Bool)
    @ OpenCL ~/.julia/packages/OpenCL/VxZWM/src/array.jl:226
  [7] unsafe_copyto!
    @ ~/.julia/packages/OpenCL/VxZWM/src/array.jl:222 [inlined]
  [8] copyto!
    @ ~/.julia/packages/OpenCL/VxZWM/src/array.jl:175 [inlined]
  [9] copyto!
    @ ~/.julia/packages/OpenCL/VxZWM/src/array.jl:181 [inlined]
 [10] CLArray{Int64, 1}(hostarray::Vector{Int64}; kwargs::@Kwargs{})
    @ OpenCL ~/.julia/packages/OpenCL/VxZWM/src/array.jl:119
 [11] CLArray
    @ ~/.julia/packages/OpenCL/VxZWM/src/array.jl:117 [inlined]
 [12] CLArray(A::Vector{Int64})
    @ OpenCL ~/.julia/packages/OpenCL/VxZWM/src/array.jl:123
 [13] top-level scope
    @ REPL[3]:1
 [14] top-level scope
    @ ~/.julia/packages/OpenCL/VxZWM/lib/cl/api.jl:113

[maleadt]

What does cl.svm_capabilities(cl.device()) return?

[VarLad]

For both my NVIDIA and AMD GPUs (with Zink), that is, running with:

RUSTICL_ENABLE=zink julia

I get: (coarse_grain_buffer = false, fine_grain_buffer = false, fine_grain_system = false)

I'm guessing that I'll need to wait for proper SVM support in RustiCL to start using OpenCL.jl.

---

That being said, this is different for CPU.
For CPU (running with RUSTICL_ENABLE=llvmpipe julia)
I get (coarse_grain_buffer = true, fine_grain_buffer = true, fine_grain_system = true)
But when I try creating an array, I get a slightly different error:

julia> CLArray([1,2])
2-element CLArray{Int64, 1}:
Error showing value of type CLArray{Int64, 1}:
ERROR: CLError(code=-59, CL_INVALID_OPERATION)
Stacktrace:
  [1] throw_api_error(res::Int32)
    @ OpenCL.cl ~/.julia/packages/OpenCL/VxZWM/lib/cl/libopencl.jl:3
  [2] check
    @ ~/.julia/packages/OpenCL/VxZWM/lib/cl/libopencl.jl:14 [inlined]
  [3] clEnqueueSVMMemcpy
    @ ~/.julia/packages/OpenCL/VxZWM/lib/cl/api.jl:34 [inlined]
  [4] enqueue_svm_memcpy(dst::Ptr{Int64}, src::Ptr{Int64}, nbytes::Int64; blocking::Bool, wait_for::Vector{OpenCL.cl.Event})
    @ OpenCL.cl ~/.julia/packages/OpenCL/VxZWM/lib/cl/svm.jl:63
  [5] enqueue_svm_memcpy
    @ ~/.julia/packages/OpenCL/VxZWM/lib/cl/svm.jl:57 [inlined]
  [6] unsafe_copyto!(dst::Vector{Int64}, dst_off::Int64, src::CLArray{Int64, 1}, src_off::Int64, N::Int64; blocking::Bool)
    @ OpenCL ~/.julia/packages/OpenCL/VxZWM/src/array.jl:226
  [7] unsafe_copyto!
    @ ~/.julia/packages/OpenCL/VxZWM/src/array.jl:222 [inlined]
  [8] copyto!
    @ ~/.julia/packages/OpenCL/VxZWM/src/array.jl:191 [inlined]
  [9] copyto!
    @ ~/.julia/packages/OpenCL/VxZWM/src/array.jl:202 [inlined]
 [10] Array
    @ ~/.julia/packages/OpenCL/VxZWM/src/array.jl:127 [inlined]
 [11] Array
    @ ./boot.jl:605 [inlined]
 [12] convert
    @ ./array.jl:607 [inlined]
 [13] adapt_storage
    @ ~/.julia/packages/GPUArrays/qt4ax/src/host/abstractarray.jl:115 [inlined]
 [14] adapt_structure
    @ ~/.julia/packages/Adapt/7T9au/src/Adapt.jl:57 [inlined]
 [15] adapt
    @ ~/.julia/packages/Adapt/7T9au/src/Adapt.jl:40 [inlined]
 [16] print_array
    @ ~/.julia/packages/GPUArrays/qt4ax/src/host/abstractarray.jl:118 [inlined]
 [17] show(io::IOContext{Base.TTY}, ::MIME{Symbol("text/plain")}, X::CLArray{Int64, 1})
    @ Base ./arrayshow.jl:399
 [18] (::REPL.var"#68#69"{REPL.REPLDisplay{REPL.LineEditREPL}, MIME{Symbol("text/plain")}, Base.RefValue{Any}})(io::Any)
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:348
 [19] with_repl_linfo(f::Any, repl::REPL.LineEditREPL)
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:646
 [20] display(d::REPL.REPLDisplay, mime::MIME{Symbol("text/plain")}, x::Any)
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:334
 [21] display
    @ ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:353 [inlined]
 [22] display(x::Any)
    @ Base.Multimedia ./multimedia.jl:340
 [23] #invokelatest#2
    @ ./essentials.jl:1054 [inlined]
 [24] invokelatest
    @ ./essentials.jl:1051 [inlined]
 [25] print_response(errio::IO, response::Any, show_value::Bool, have_color::Bool, specialdisplay::Union{Nothing, AbstractDisplay})
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:390
 [26] (::REPL.var"#70#71"{REPL.LineEditREPL, Pair{Any, Bool}, Bool, Bool})(io::Any)
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:359
 [27] with_repl_linfo(f::Any, repl::REPL.LineEditREPL)
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:646
 [28] print_response(repl::REPL.AbstractREPL, response::Any, show_value::Bool, have_color::Bool)
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:357
 [29] (::REPL.var"#do_respond#96"{Bool, Bool, REPL.var"#112#130"{…}, REPL.LineEditREPL, REPL.LineEdit.Prompt})(s::REPL.LineEdit.MIState, buf::Any, ok::Bool)
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:988
 [30] #invokelatest#2
    @ ./essentials.jl:1054 [inlined]
 [31] invokelatest
    @ ./essentials.jl:1051 [inlined]
 [32] run_interface(terminal::REPL.Terminals.TextTerminal, m::REPL.LineEdit.ModalInterface, s::REPL.LineEdit.MIState)
    @ REPL.LineEdit ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/LineEdit.jl:2749
 [33] run_frontend(repl::REPL.LineEditREPL, backend::REPL.REPLBackendRef)
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:1456
 [34] (::REPL.var"#75#81"{REPL.LineEditREPL, REPL.REPLBackendRef})()
    @ REPL ~/.julia/juliaup/julia-1.11.0-rc4+0.x64.linux.gnu/share/julia/stdlib/v1.11/REPL/src/REPL.jl:461
Some type information was truncated. Use `show(err)` to see complete types.

[VarLad]

Is the latter a display issue where its having trouble displaying the elements of type CLAarray on the REPL? 🤔

[maleadt]

Is the latter a display issue where its having trouble displaying the elements of type CLAarray on the REPL? 🤔

Looks like it. Does manually copying back to an Array work?

I get: (coarse_grain_buffer = false, fine_grain_buffer = false, fine_grain_system = false)

Guess I didn't add the check; the CLArray constructor should probably detect this and error out.

[VarLad]

Apparently this works (for CPU):

julia> x = CLArray([1,2]);

julia> y = CLArray([2,3]);

julia> v = x + y;

julia> u = x .* y;

julia> typeof(u)
CLArray{Int64, 1}

[maleadt]

That's interesting. I wonder if it's something to do with the cross-queue usage of the allocations (the REPL display code runs on another task, which uses another queue).

[VarLad]

I believe the current issue with this after the latest release is with implementation of coarse-grained SVM (or any of the other supported allocation strategies) in Mesa itself.
Hopefully that'll get fixed by https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32942

I'll reopen the issue if there're still issues after that PR lands.