Accidentally calling GPU intrinsics on the host causes segfaults

[Keno]

CUDA_VISIBLE_DEVICES=3 julia --project=..
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.0.3 (2018-12-18)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org/ release
|__/                   |

julia> using CUDAnative

julia> f(x) = CUDAnative.tanh(x)
f (generic function with 1 method)

julia> @code_typed f(1f0)
CodeInfo(
1 1 ─ %1 = Base.llvmcall::IntrinsicFunction                                                                                              │╻╷ tanh
  │   %2 = (%1)(("declare float @__nv_tanhf(float)", "%2 =  call float @__nv_tanhf(float %0)\nret float %2"), Float32, Tuple{Float32}, x)::Float32o expansion
  └──      return %2                                                                                                                     │
) => Float32

julia> f(1f0)

signal (11): Segmentation fault
in expression starting at no file:0
_ITM_registerTMCloneTable at /opt/julia/julia-1.0.3/bin/../lib/julia/libspqr.so (unknown line)
jl_apply_generic at /buildworker/worker/package_linux64/build/src/gf.c:2184
eval_user_input at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.0/REPL/src/REPL.jl:82
macro expansion at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.0/REPL/src/REPL.jl:117 [inlined]
JuliaGPU/CUDAnative.jl#28 at ./task.jl:259
jl_apply_generic at /buildworker/worker/package_linux64/build/src/gf.c:2184
jl_apply at /buildworker/worker/package_linux64/build/src/julia.h:1537 [inlined]
start_task at /buildworker/worker/package_linux64/build/src/task.c:268
unknown function (ip: 0xffffffffffffffff)
Allocations: 4184659 (Pool: 4183753; Big: 906); GC: 8
Segmentation fault (core dumped)

My suggestion would be to

1. Switch this to ccall with llvmcall calling convention
2. Add an llvmcall_gpu calling convention that emits an error when somebody tries to compile it for a host backend.

cc @maleadt

[maleadt]

• Switch this to ccall with llvmcall calling convention

• Add an llvmcall_gpu calling convention that emits an error when somebody tries to compile it for a host backend.

These functions aren't proper intrinsics, so ccall refuses to call them.
https://github.com/JuliaLang/julia/blob/69d4ee80432e999e340966e90bc48d32fd5c1cb8/src/ccall.cpp#L1953-L1954

Beyond that, Base currently doesn't know we're generating GPU code, so where would propose to add this distinction? We currently just jl_get_llvmf_defn the function and hack its triple etc during the module_setup hook. So I've been trying to avoid adding a GPU-mode or other custom IRgen to Base...

I had imagined implementing these intrinsics use Cassette and not have them generate GPU IR outside of the GPU context (and for custom functionality make sure there's a fallback CPU context that just errors), but Cassette usability for GPUs has been progressing more slowly than anticipated.