ERROR_ILLEGAL_ADDRESS when using ternary operator in map! on CuArray

[jamccauley3]

When using map! with a conditional on a CuArray there is an error when using a ternary operator but not when multiplying by the result of the conditional on the most recent CUDA.jl version. There are also no issues using map! with these inputs without the conditional. Below are the three implementations of map! I've used. I wasn't able to replicate this issue on a simplified version of this code, which I've also included. Any help is appreciated, thanks.

Setup for the following functions:
All variables are CUDA variables

lnA, Q, n, R = dynamics.lnA, dynamics.E, dynamics.n, dynamics.R
dt = Δt(dynamics, uₖ)
A_diag = @view A[1:(size(A, 1)+1):end]

map! without conditional (no errors):

map!((yₖ, Tₖ) -> 1.0 / (1.0 + exp(lnA - Q / R / Tₖ - yₖ / n) * dt), A_diag, xₖ, uₖ)

map! with conditional and ternary operator:
Error: CuError(CUDA.cudaError_enum(0x000002bc))
Error: illegal memory access was encountered (code 700, ERROR_ILLEGAL_ADDRESS)

map!((yₖ, Tₖ) -> Tₖ > 1.0 ? 0.0f0 : 1.0 / (1.0 + exp(lnA - Q / R / Tₖ - yₖ / n) * dt), A_diag, xₖ, uₖ)

map! multiplying by conditional (no errors):

map!((yₖ, Tₖ) -> (Tₖ < 1.0) * 1.0 / (1.0 + exp(lnA - Q / R / Tₖ - yₖ / n) * dt), A_diag, xₖ, uₖ)

I was unable to replicate this issue with a simplified version of this code, so I included my full implementation. Below is a simplified version of my code that works fine for both conditional map! implementations:

using CUDA

A = cu(ones(50))
B = cu(1:50)
C = cu(zeros(50, 100))
D = cu(zeros(50, 100))
C_diag = @view C[1:(size(C, 1)+1):end]
D_diag = @view D[1:(size(D, 1)+1):end]


map!((a, b) -> (b < 25.0) * a, C_diag, A, B)

@show C_diag

map!((a, b) -> b > 25.0 ? 0 : a, D_diag, A, B)

@show D_diag

Details on Julia:

Julia Version 1.11.4
Commit 8561cc3d68d (2025-03-10 11:36 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 32 × AMD Ryzen Threadripper PRO 5955WX 16-Cores
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, znver3)
Threads: 1 default, 0 interactive, 1 GC (on 32 virtual cores)

Details on CUDA:

CUDA runtime 12.8, artifact installation
CUDA driver 12.6
NVIDIA driver 560.35.3

CUDA libraries: 
- CUBLAS: 12.8.4
- CURAND: 10.3.9
- CUFFT: 11.3.3
- CUSOLVER: 11.7.3
- CUSPARSE: 12.5.8
- CUPTI: 2025.1.1 (API 26.0.0)
- NVML: 12.0.0+560.35.3

Julia packages: 
- CUDA: 5.7.2
- CUDA_Driver_jll: 0.12.1+1
- CUDA_Runtime_jll: 0.16.1+0

Toolchain:
- Julia: 1.11.4
- LLVM: 16.0.6

Preferences:
- default_memory: device

2 devices:
  0: NVIDIA GeForce RTX 4090 (sm_89, 3.061 GiB / 23.988 GiB available)
  1: NVIDIA GeForce RTX 4090 (sm_89, 22.936 GiB / 23.988 GiB available)

[maleadt]

I'd guess you're accidentally referencing a global CPU variable, leading to the GPU reading inaccessible memory. We currently don't spot those during compilation: JuliaGPU/GPUCompiler.jl#232

In any case, without a functioning reproducer there's not much we can do about this.

[jamccauley3]

Why would using a ternary operator reference a global CPU variable but just a conditional statement does not? That's the only difference between the working and not-working lines of code, which is confusing to me.

[maleadt]

It may influence optimization and how things get inlined or not.

In any case, can you share the @device_code output of the failing example?