Correct MLIR pass to compile `test/Integration/GPU/CUDA/async.mlir`

MLIR
1

I was trying to compile the examples in https://github.com/llvm/llvm-project/tree/main/mlir/test/Integration/GPU/CUDA to executables. I can compile all of them except async.mlir.

My passes for async.mlir are like this:

mlir-opt async.mlir \
    --gpu-kernel-outlining \
    --pass-pipeline='gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin)' \
    --gpu-async-region \
    --gpu-to-llvm \
    --async-to-async-runtime \
    --async-runtime-ref-counting \
    --convert-async-to-llvm \
    --convert-std-to-llvm 2>&1 >bin/async.mlir.out

mlir-translate bin/async.mlir.out --mlir-to-llvmir \
    | opt -O3 -S | llc -O3 | as - -o bin/async.o

clang++-11 bin/async.o -lcuda \
    $HOME/opt/llvm/lib/libmlir_cuda_runtime.so \
    $HOME/opt/llvm/lib/libmlir_runner_utils.so \
    $HOME/opt/llvm/lib/libmlir_c_runner_utils.so \
    -o bin/async

When I run .\async, I’m getting this error –

PLEASE submit a bug report to https://bugs.llvm.org/ and include the crash backtrace.
Stack dump:
0.      Program arguments: llc -O3 
1.      Running pass 'Function Pass Manager' on module '<stdin>'.
2.      Running pass 'X86 DAG->DAG Instruction Selection' on function '@async_execute_fn'
#0 0x00007f3653e1842f llvm::sys::PrintStackTrace(llvm::raw_ostream&) (/usr/lib/x86_64-linux-gnu/libLLVM-11.so.1+0xaa642f)
#1 0x00007f3653e16790 llvm::sys::RunSignalHandlers() (/usr/lib/x86_64-linux-gnu/libLLVM-11.so.1+0xaa4790)
#2 0x00007f3653e18905 (/usr/lib/x86_64-linux-gnu/libLLVM-11.so.1+0xaa6905)
#3 0x00007f365334a3c0 __restore_rt (/lib/x86_64-linux-gnu/libpthread.so.0+0x153c0)
#4 0x00007f365437c730 (/usr/lib/x86_64-linux-gnu/libLLVM-11.so.1+0x100a730)
#5 0x00007f3654385ee8 (/usr/lib/x86_64-linux-gnu/libLLVM-11.so.1+0x1013ee8)
#6 0x00007f3654387331 (/usr/lib/x86_64-linux-gnu/libLLVM-11.so.1+0x1015331)
#7 0x0000000002177e02 
{standard input}: Assembler messages:
{standard input}: Warning: end of file not at end of a line; newline inserted
{standard input}:140: Error: bad register name `%rd'
{standard input}: Error: open CFI at the end of file; missing .cfi_endproc directive
clang: error: no such file or directory: 'bin/async.o'

However, I can run using mlir-cpu-runner without compiling, like this –

mlir-opt async.mlir \
    -gpu-kernel-outlining \
    -pass-pipeline='gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin)' \
    -gpu-async-region \
    -gpu-to-llvm \
    -async-to-async-runtime \
    -async-runtime-ref-counting \
    -convert-async-to-llvm \
    -convert-std-to-llvm \
| mlir-cpu-runner \
    --shared-libs=/$HOME/opt/llvm/lib/libmlir_cuda_runtime.so \
    --shared-libs=/$HOME/opt/llvm/lib/libmlir_async_runtime.so \
    --shared-libs=/$HOME/opt/llvm/lib/libmlir_runner_utils.so \
    --entry-point-result=void -O0

and there is no problem, it runs.

But how do I compile it to executable to run on GPU?

2

Hello @chudur-budur
Did you find way to make it work?

3

No, unfortunately no solution yet.

4

Hello @chudur-budur

I have posted the similar question and found the solution. But important thing to mention is that, this solution is for LLVM of version 18.

You can find the discussion under this thread How to generate nvidia cuda bin (cubin) from MLIR?

Hope it helps :slightly_smiling_face:

Thanks!