`sind` in kernels not supported / causes errors

[maximilian-gelbrecht]

I can see in the repo that there's some support for trig functions like sind but they currently don't work in kernels:

using Reactant, KernelAbstractions, CUDA

# get backend 
ReactantKernelAbstractionsExt = Base.get_extension(
    Reactant, :ReactantKernelAbstractionsExt
)
ReactantBackend = ReactantKernelAbstractionsExt.ReactantBackend()

# Simple kernel that uses sind
@kernel function sind_kernel!(out, x)
    i = @index(Global)
    out[i] = sind(x[i]) # this causes an error
    #out[i] = sin(rad2deg(x[i])) # this would work
end

# Create input array
x_cpu = Float32[0, 30, 45, 60, 90]
x_reactant = Reactant.to_rarray(x_cpu)
out_reactant = Reactant.to_rarray(zeros(Float32, 5))

# Launch kernel
kernel! = sind_kernel!(ReactantBackend)
@jit kernel!(out_reactant, x_reactant; ndrange=5)

[10944] signal 4 (1): Illegal instruction: 4
in expression starting at REPL[9]:1
Allocations: 71344590 (Pool: 71343371; Big: 1219); GC: 39
unknown function (ip: 0x1052702c4)
Allocations: 71344590 (Pool: 71343371; Big: 1219); GC: 39
Trace/BPT trap: 5

[wsmoses]

can you upload the pre_all mlir ?

[maximilian-gelbrecht]

Do you mean the output of

mlir = @code_hlo optimize=false kernel!(out_reactant, x_reactant; ndrange=5)

with that?

I pasted it here on gist

[wsmoses]

nope, not quite set

export JULIA_DEBUG="Reactant" [in the env before running julia]

then set Reactant.MLIR.IR.DUMP_MLIR_ALWAYS[] = true

before running your program.

adding a pr on this to the docs would be super helpful! [cc @avik-pal @giordano ]

[maximilian-gelbrecht]

Ok, here it is

https://gist.github.com/maximilian-gelbrecht/b84b0be59c528771e09f7e43332a9ee9

[wsmoses]

Okay I see what's up. It's fixable but it's going to be nontrivial.

Leave this open, but can you make a workaround for now?

[maximilian-gelbrecht]

Yes sure, I already adjusted it to a simple sin(deg2rad(... )) in my code.