This PR adds device existence validation to GPU operators to prevent cryptic failures when CUDA_VISIBLE_DEVICES and container GPU configuration are mismatched.

Problem

When using Docker with both --env CUDA_VISIBLE_DEVICES=1 and --gpus "device=1", the container runtime only exposes device 0 but renumbers it from the original device 1. This causes Devito to attempt using device ID 1 (from CUDA_VISIBLE_DEVICES) on a system that only has device 0 available, leading to opaque failures like:

tests/test_gpu_openacc.py ...............
Error: Process completed with exit code 1.

Solution

Added device validation logic that generates runtime checks before calling acc_set_device_num() or omp_set_default_device(). The generated code now includes:

OpenACC validation:

if (deviceid != -1) {
    int ngpus = acc_get_num_devices(acc_device_nvidia);
    if (deviceid >= ngpus) {
        printf("OpenACC: Error - device %d >= %d devices\n", deviceid, ngpus);
        exit(1);
    }
    acc_set_device_num(deviceid, acc_device_nvidia);
}

OpenMP validation:

if (deviceid != -1) {
    int ngpus = omp_get_num_devices();
    if (deviceid >= ngpus) {
        printf("OpenMP: Error - device %d >= %d devices\n", deviceid, ngpus);
        exit(1);
    }
    omp_set_default_device(deviceid);
}

Benefits

• Clear error messages: Users get specific error messages instead of cryptic exit codes
• Graceful failure: Program exits cleanly with error code 1 instead of undefined behavior
• Backward compatible: Only validates when deviceid is explicitly set (deviceid != -1)
• Cross-platform: Works for both OpenACC and OpenMP backends
• Minimal overhead: Uses existing device count infrastructure with concise validation

Implementation Details

The solution follows Devito's established patterns:

1. Leveraged existing infrastructure: Uses the existing 'num-devices' mappers for generating appropriate runtime calls (omp_get_num_devices or acc_get_num_devices)

2. Minimal code changes: Enhanced device setup paths to include validation logic before device assignment, following Devito's 90-character line limit and coding standards

3. Proper integration: Works seamlessly with both sequential and MPI device setup configurations

Changes

• Modified devito/passes/iet/langbase.py to add concise validation logic using established patterns
• Updated tests in tests/test_gpu_openmp.py and tests/test_gpu_openacc.py to validate the new behavior
• Follows Devito's coding standards with proper imports, line limits, and minimal implementation

Fixes #2711.

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.