On Arm platform tests for group functions failed when targeting omp.accelerated

[ouankou]

Bug summary
Acpp can be built without issue on Oracle A1 VPS (4-core Ampere Altra chip). However, the built-in tests for group functions all failed.
I saw this issue related to Arm: #1044
It seems acpp has been tested on Arm. I'm wondering if I didn't set up acpp correctly.

To Reproduce

# build acpp and builtin tests
./sycl_tests --run_test=group_functions_tests

Expected behavior

All tests should pass.

Describe your setup

CPU: Ampere Altra (Arm Neoverse-N1, 4 cores)
RAM: 24 GB
OS: Debian 12
Software: Boost 1.74, Python 3.11

• how you have installed AdaptiveCpp, and which version/branch/git commit:
  Build from source in GitHub repo develop branch
• Describe the dependencies that AdaptiveCpp sits on top of in your setup:
  • clang version and how clang was installed:
    Clang/LLVM 17.0.6 from official LLVM apt repo.
  • host compiler:
    GCC 12.2 from official Debian repo
  • CUDA version (if applicable)
  • ROCm version and how ROCm was installed (if applicable)
• How you have compiled your application and which arguments you have passed to acpp. In particular, which backends and hardware you have compiled for:

cmake -DWITH_SSCP_COMPILER=OFF -DWITH_ACCELERATED_CPU=ON -DWITH_OPENCL_BACKEND=OFF -DCMAKE_INSTALL_PREFIX=$HOME/Projects/AdaptiveCpp/install -DWITH_CUDA_BACKEND=OFF -DWITH_ROCM_BACKEND=OFF -DWITH_LEVEL_ZERO_BACKEND=OFF -DDEFAULT_TARGETS='omp.accelerated' ..

Optional additional diagnostic information
The following information can potentially help us:

• For compilation/toolchain or setup-related issues: Setting the environment variable HIPSYCL_DRYRUN=1 during compilation will cause syclcc to print the commands it would execute instead of actually executing. This can help verify the sanity of your setup.
• syclcc --hipsycl-version prints information about the AdaptiveCpp configuration that may be interesting for setup-related issues.

acpp warning: syclcc is deprecated; please use acpp instead.
acpp [AdaptiveCpp compilation driver], Copyright (C) 2018-2023 Aksel Alpay and the AdaptiveCpp project
  AdaptiveCpp version: 24.02.0+git.0359cac9.20240401.branch.develop.dirty
  Installation root: /home/ouankou/Projects/AdaptiveCpp/install
  Plugin LLVM version: 17, can accelerate CPU: True
  Available runtime backends:
     librt-backend-omp.so


Full configuration [can be overridden using environment variables or command line arguments]:
    version-major: 24
    version-minor: 02
    version-patch: 0
    version-suffix: +git.0359cac9.20240401.branch.develop.dirty
    plugin-llvm-version-major: 17
    plugin-with-cpu-acceleration: true
    default-clang: /usr/lib/llvm-17/bin/clang++
    default-targets: omp.accelerated
    default-cpu-cxx: /usr/bin/c++
    default-rocm-path: /opt/rocm
    default-use-bootstrap-mode: false
    default-is-dryrun: false
    default-use-accelerated-cpu: true
    default-clang-include-path: /usr/lib/llvm-17/lib/clang/17/include/..
    default-sequential-link-line: -L/usr/lib/aarch64-linux-gnu -lboost_context -lboost_fiber -Wl,-rpath=/usr/lib/aarch64-linux-gnu
    default-sequential-cxx-flags: -I/usr/include -D_ENABLE_EXTENDED_ALIGNED_STORAGE
    default-omp-link-line: -L/usr/lib/aarch64-linux-gnu -lboost_context -lboost_fiber -Wl,-rpath=/usr/lib/aarch64-linux-gnu -fopenmp
    default-omp-cxx-flags: -I/usr/include -fopenmp -D_ENABLE_EXTENDED_ALIGNED_STORAGE
    default-is-explicit-multipass: false
    default-save-temps: false

• For issues related to the runtime, setting the environment variable ACPP_DEBUG_LEVEL=3 will cause it to print a lot of diagnostic information that might be helpful. Attach the output of your program with ACPP_DEBUG_LEVEL=3 if you think it might be helpful for your issue.
  Execution output (fail.log) and DEBUG output(fail_output.log) are attached.
  fail.log
  fail_debug.log

• Recent AdaptiveCpp versions include a tool called acpp-info, which will print information about available backends and devices. This may be interesting for issues related to e.g. device visibility.

=================Backend information===================
Loaded backend 0: OpenMP
  Found device: hipSYCL OpenMP host device

=================Device information===================
***************** Devices for backend OpenMP *****************
Device 0:
 General device information:
  Name: hipSYCL OpenMP host device
  Backend: OpenMP
  Vendor: the hipSYCL project
  Arch: <native-cpu>
  Driver version: 1.2
  Is CPU: 1
  Is GPU: 0
 Default executor information:
  Is in-order queue: 0
  Is out-of-order queue: 1
  Is task graph: 0
 Device support queries:
  images: 0
  error_correction: 0
  host_unified_memory: 1
  little_endian: 1
  global_mem_cache: 1
  global_mem_cache_read_only: 0
  global_mem_cache_read_write: 1
  emulated_local_memory: 1
  sub_group_independent_forward_progress: 0
  usm_device_allocations: 1
  usm_host_allocations: 1
  usm_atomic_host_allocations: 1
  usm_shared_allocations: 1
  usm_atomic_shared_allocations: 1
  usm_system_allocations: 1
  execution_timestamps: 1
  sscp_kernels: 0
 Device properties:
  max_compute_units: 4
  max_global_size0: 18446744073709551615
  max_global_size1: 18446744073709551615
  max_global_size2: 18446744073709551615
  max_group_size: 1024
  max_num_sub_groups: 18446744073709551615
  preferred_vector_width_char: 4
  preferred_vector_width_double: 1
  preferred_vector_width_float: 1
  preferred_vector_width_half: 2
  preferred_vector_width_int: 1
  preferred_vector_width_long: 1
  preferred_vector_width_short: 2
  native_vector_width_char: 4
  native_vector_width_double: 1
  native_vector_width_float: 1
  native_vector_width_half: 2
  native_vector_width_int: 1
  native_vector_width_long: 1
  native_vector_width_short: 2
  max_clock_speed: 0
  max_malloc_size: 18446744073709551615
  address_bits: 64
  max_read_image_args: 0
  max_write_image_args: 0
  image2d_max_width: 0
  image2d_max_height: 0
  image3d_max_width: 0
  image3d_max_height: 0
  image3d_max_depth: 0
  image_max_buffer_size: 0
  image_max_array_size: 0
  max_samplers: 0
  max_parameter_size: 18446744073709551615
  mem_base_addr_align: 8
  global_mem_cache_line_size: 64
  global_mem_cache_size: 1
  global_mem_size: 18446744073709551615
  max_constant_buffer_size: 18446744073709551615
  max_constant_args: 18446744073709551615
  local_mem_size: 18446744073709551615
  printf_buffer_size: 18446744073709551615
  partition_max_sub_devices: 0
  vendor_id: 18446744073709551615
  sub_group_sizes: 1

Additional context
Add any other context about the problem here.

[illuhad]

Yes, it is expected to work on ARM, and was tested on A64fx and ThunderX2, but I don't know if anybody has ever tried it on Altra.

@nilsfriess, I recall you recently reported to have found some bugs in the tests. Were the group tests by chance among the problematic test cases?

[ouankou]

Thanks for the info!

I've also tested it on the Apple M1 Arm chip. Except for acpp, all dependencies are installed from official Debian or LLVM repos.
System: Debian 12 in the UTM virtual machine
Software: LLVM 15/16/17, Boost 1.74.
Building config:

cmake -DWITH_SSCP_COMPILER=OFF -DWITH_ACCELERATED_CPU=ON -DWITH_OPENCL_BACKEND=OFF -DCMAKE_INSTALL_PREFIX=$HOME/Projects/AdaptiveCpp/install -DWITH_CUDA_BACKEND=OFF -DWITH_ROCM_BACKEND=OFF -DWITH_LEVEL_ZERO_BACKEND=OFF -DDEFAULT_TARGETS='omp.accelerated' ..

Identical failures to group functions happened.

[nilsfriess]

@nilsfriess, I recall you recently reported to have found some bugs in the tests. Were the group tests by chance among the problematic test cases?

I only found some UB in the tests, but I don't think the group tests were among the problematic tests. But I'll check again

[nilsfriess]

I only found some UB in the tests, but I don't think the group tests were among the problematic tests. But I'll check again

Checked again, there is actually a problem in the group tests (an integer overflow, to be specific). Not sure if this can explain the error reported above.

[illuhad]

@ouankou Can you check whether it works as expected with omp.library-only target?