NVVM IR Invalid data layout when building with Cuda 12.0

[jorge-ortega]

When using CUDA 12.0, verifying the NVVR IR generated fails with the following error:

  Error: Invalid data layout :
  Error: e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64
  Error: 32-bit: e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64
  Error: 64-bit: e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64

The format is deprecated but still works with 12.1 and higher. This might be a bug in the verifier in 12.0 that we can ignore. This would be best fixed by using the supported data layout when using CUDA 12 and up.

[tlaakkonen]

I'm getting the same error with Cuda 12.8. But I saw on another thread you were able to work with 12.8 - is there a workaround you're using?

Output of cargo build:

...
 thread 'rustc' panicked at /home/epyc/.cargo/git/checkouts/rust-cuda-f2e68423bdc84998/290d711/crates/rustc_codegen_nvvm/src/nvvm.rs:120:9:
  Malformed NVVM IR program rejected by libnvvm, dumping verifier log:

  Error: Invalid data layout :
  Error: e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64
  Error: 32-bit: e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64
  Error: 64-bit: e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64
...

Output of nvcc --version:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Fri_Feb_21_20:23:50_PST_2025
Cuda compilation tools, release 12.8, V12.8.93
Build cuda_12.8.r12.8/compiler.35583870_0

Output of $LLVM_CONFIG --version: 7.0.1

Edit: Fixed, there was an old version of Cuda 12.0 hanging around that needed to be purged properly.

[jorge-ortega]

Haven't needed any workarounds for 12.8 as of yet. I've been using LLVM 7.1.0, but unsure if that would make a difference. If you could share the IR, it could help further. I think all you need to do is call final_module_path on your cuda builder with a path to where the IR should go.

I also have a patch I've been sitting on that updates the data layout. I can push that to a branch if you'd like to try that.

[brandonros]

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Fri_Jan__6_16:45:21_PST_2023
Cuda compilation tools, release 12.0, V12.0.140
Build cuda_12.0.r12.0/compiler.32267302_0

$ cat Cargo.toml
[package]
name = "ed25519_vanity"

As per https://github.com/Rust-GPU/Rust-CUDA/blob/290d711507d26f3ca16a484ba02c8a03ba9d7b7c/guide/src/nvvm/technical/debugging.md#miscompilations

Use RUSTFLAGS="--emit=llvm-ir" and find crate_name.ll in target/nvptx64-nvidia-cuda/<debug/release>/deps/ and attach it in any bug report.

RUSTFLAGS="--emit=llvm-ir" cargo build

find target -type f | grep ".ll"

I do not have anything other than target/debug/build/ed25519_vanity-64976b22a8e88420/build_script_build-64976b22a8e88420.ll that looks close

Edit:

I think all you need to do is call final_module_path on your cuda builder with a path to where the IR should go.

Got it.

use std::env;
use std::path;

use cuda_builder::CudaBuilder;

fn main() {
    println!("cargo:rerun-if-changed=build.rs");
    println!("cargo:rerun-if-changed=kernels");

    let out_path = path::PathBuf::from(env::var("OUT_DIR").unwrap());
    let builder = CudaBuilder::new("kernels")
        .copy_to(out_path.join("kernels.ptx"))
        .final_module_path(out_path.join("kernels.ll"))
        .build()
        .unwrap();
}

generates target/debug/build/ed25519_vanity-f0fe2ffa5f8b0bad/out/kernels.ll

vecadd example:

use cuda_std::prelude::*;

#[kernel]
#[allow(improper_ctypes_definitions, clippy::missing_safety_doc)]
pub unsafe fn vecadd(a: &[f32], b: &[f32], c: *mut f32) {
    let idx = thread::index_1d() as usize;
    if idx < a.len() {
        let elem = unsafe { &mut *c.add(idx) };
        *elem = a[idx] + b[idx];
    }
}

; ModuleID = 'merged_modules'
source_filename = "merged_modules"
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
target triple = "nvptx64-nvidia-cuda"

; Function Attrs: nounwind
define void @vecadd([0 x float]* nocapture readonly, i64, [0 x float]* nocapture readonly, i64, float* nocapture) unnamed_addr #0 {
  %6 = tail call i32 @__nvvm_thread_idx_x() #0
  %7 = icmp ult i32 %6, 1025
  tail call void @llvm.assume(i1 %7) #0
  %8 = tail call i32 @__nvvm_block_idx_x() #0
  %9 = icmp sgt i32 %8, -1
  tail call void @llvm.assume(i1 %9) #0
  %10 = tail call i32 @__nvvm_block_dim_x() #0
  %11 = icmp ne i32 %10, 0
  tail call void @llvm.assume(i1 %11) #0
  %12 = icmp ult i32 %10, 1026
  tail call void @llvm.assume(i1 %12) #0
  %13 = mul i32 %10, %8
  %14 = add i32 %13, %6
  %15 = zext i32 %14 to i64
  %16 = icmp ult i64 %15, %1
  br i1 %16, label %18, label %17

; <label>:17:                                     ; preds = %20, %5
  ret void

; <label>:18:                                     ; preds = %5
  %19 = icmp ult i64 %15, %3
  br i1 %19, label %20, label %27

; <label>:20:                                     ; preds = %18
  %21 = getelementptr inbounds [0 x float], [0 x float]* %0, i64 0, i64 %15
  %22 = load float, float* %21, align 4
  %23 = getelementptr inbounds float, float* %4, i64 %15
  %24 = getelementptr inbounds [0 x float], [0 x float]* %2, i64 0, i64 %15
  %25 = load float, float* %24, align 4
  %26 = fadd float %22, %25
  store float %26, float* %23, align 4
  br label %17

; <label>:27:                                     ; preds = %18
  tail call void @llvm.trap() #0
  unreachable
}

; Function Attrs: nounwind
declare i32 @__nvvm_thread_idx_x() unnamed_addr #0

; Function Attrs: nounwind
declare void @llvm.assume(i1) #0

; Function Attrs: nounwind
declare i32 @__nvvm_block_idx_x() unnamed_addr #0

; Function Attrs: nounwind
declare i32 @__nvvm_block_dim_x() unnamed_addr #0

; Function Attrs: noreturn nounwind
declare void @llvm.trap() #1

attributes #0 = { nounwind }
attributes #1 = { noreturn nounwind }

!nvvm.annotations = !{!0}
!nvvmir.version = !{!1}

!0 = !{void ([0 x float]*, i64, [0 x float]*, i64, float*)* @vecadd, !"kernel", i32 1}
!1 = !{i32 2, i32 0, i32 3, i32 1}

Edit 2: Confirmed not an issue / "fixed" by not using 12.0 but using 12.8 instead even with llvm 7.0.1

apt update
apt-get remove --purge 'cuda-12-0*' 'cuda-*12-0*' 'libcublas-12-0' 'libcublas-dev-12-0' cuda-12.0
apt install pkg-config libssl-dev llvm-7 llvm-7-dev llvm-7-tools clang-7 zlib1g-dev cuda-12-8