Repository navigation
Description
Problem Description
hipify-perl and hipify-clang pass #include <cuda/barrier> and #include <cuda/pipeline> through with no diagnostic. libhipcxx does not provide a working version of those headers. In amd-develop of ROCm/libhipcxx (24a4802), include/cuda/barrier stops with #error barrier is not supported on AMD hardware and should not be included under __HIP_PLATFORM_AMD__, and cuda/pipeline includes cuda/barrier. cuda/semaphore and cuda/ptx have the same kind of stub, and some other public libcu++ headers are absent from libhipcxx. So the conversion looks clean and the failure only shows up later at compile time.
Input barrier_in.cu:
#include <cuda/barrier>
#include <cooperative_groups.h>
__global__ void sync_kernel(int *out) {
__shared__ cuda::barrier<cuda::thread_scope_block> bar;
if (threadIdx.x == 0) {
init(&bar, blockDim.x);
}
__syncthreads();
out[threadIdx.x] = threadIdx.x;
bar.arrive_and_wait();
}
On amd-develop at 9a85980, hipify-perl prints no warning for this file, and for pipeline_in.cu (#include <cuda/pipeline>) and for the public CUDA samples simpleAWBarrier.cu and bf16TensorCoreGemm.cu. hipify-clang prints no header warning for pipeline_in.cu and bf16TensorCoreGemm.cu. The #include lines stay as they are, which is expected, since libhipcxx uses the same paths.
Compiling the hipify-perl output against libhipcxx:
$ hipcc -I<libhipcxx>/include --offload-arch=gfx942 --cuda-device-only -c barrier_in.hip -o /dev/null
<libhipcxx>/include/cuda/barrier:38:2: error: barrier is not supported on AMD hardware and should not be included
pipeline_in, simpleAWBarrier and bf16TensorCoreGemm fail with the same error at cuda/barrier:38:2.
Control: a file that includes only cuda/std/array and cuda/std/type_traits converts with no warning and compiles against the same libhipcxx (rc 0). So the include line itself is fine for headers libhipcxx provides.
Possible fix: print a warning at conversion time for includes of libcu++ headers that libhipcxx does not provide.
Operating System
Ubuntu 24.04 LTS (Noble Numbat)
CPU
Intel(R) Xeon(R) Gold 6130 CPU @ 2.10GHz
ROCm Version
ROCm 7.1.0 hipcc (HIP 7.1.25424). The ROCm version list in the template stops at 6.3.1.
ROCm Component
HIPIFY
Steps to Reproduce
- Build hipify-clang from
amd-develop(9a85980) with LLVM 20.1.8 and CUDA 12.8, and clone ROCm/libhipcxx (amd-develop, 24a4802). hipify-perl barrier_in.cu > barrier_in.hiphipcc -I<libhipcxx>/include --offload-arch=gfx942 --cuda-device-only -c barrier_in.hip -o /dev/null
Additional Information
No GPU is needed, only compilation. hipify-clang itself fails on barrier_in.cu and simpleAWBarrier.cu with error: initialization is not supported for __shared__ variables when it parses the cuda::barrier shared variable, before and after the change that adds the warning, so the hipify-clang half of this report uses pipeline_in.cu and bf16TensorCoreGemm.cu.
Activity