First, let's break down the root cause of your error. The truncated error message showing -cudar (instead of -cudart=shared) and the messed-up command line hints that your nvcc parameters are being split incorrectly by the cc crate.

Key Issues in Your Current build.rs #

1. Split -gencode parameter: NVCC requires -gencode arch=compute_61,code=sm_61 to be passed as a single command-line argument, but you're splitting it into two separate .flag() calls. This confuses nvcc's argument parser, leading to malformed commands.
2. Misaligned argument parsing: The broken parameter handling cascades into other parts of the command line, causing fragments like -cudar (a truncated -cudart=shared) to appear.

Corrected build.rs Implementation #

Here's the fixed version with proper parameter handling:

extern crate cc;

fn main() {
    let mut build = cc::Build::new();
    
    // Enable CUDA compilation mode
    build.cuda(true);
    
    // Pass -cudart=shared as a single unbroken flag
    build.flag("-cudart=shared");
    
    // Critical: Pass the entire -gencode argument as one flag (don't split!)
    build.flag("-gencode arch=compute_61,code=sm_61");
    
    // Add your CUDA kernel source file
    build.file("kernel/kernel.cu");
    
    // Compile to a static library for Rust to link against
    build.compile("kernel.a");
}

Additional Verification & Adjustment Tips #

• Inspect the full command line: Run cargo build -vv to see the complete nvcc command being executed. Verify that -cudart=shared and -gencode arch=compute_61,code=sm_61 appear as single, unbroken arguments.
• Tweak optimization levels: If you need to override default settings (e.g., for Release builds), add .opt_level(3) to the build chain to enable maximum optimization.
• Confirm CUDA toolkit accessibility: Ensure nvcc is in your system PATH, or set the CUDA_PATH environment variable if your toolkit is installed in a non-standard location.

Why This Works #

• NVCC expects -gencode to be followed immediately by its architecture specification as part of the same argument. Splitting them makes nvcc interpret -gencode as an incomplete option, triggering parsing errors that mess up subsequent parameters.
• Keeping -cudart=shared as a single flag ensures nvcc correctly recognizes the CUDA runtime linkage mode without splitting the argument at the equals sign.

内容的提问来源于stack exchange，提问作者Ltei