Limitations and unsupported APIs#

2026-10-23

3 min read time

Applies to Linux and Windows

Platform limitations#

libhipcxx has the following platform limitations.

  • libhipcxx does not support the CUDA backend or NVIDIA hardware.

  • cuda::std::chrono::system_clock::now() does not return a UNIX timestamp. The host system clock and the device system clock are not synchronized and may run at different clock rates.

Important

Some APIs from libcudacxx may be missing or may not achieve the same performance on AMD GPUs due to fundamental hardware architecture differences between AMD and NVIDIA GPUs. APIs that rely on NVIDIA-specific hardware features (such as PTX instructions, tensor cores, or specific compute capabilities) may be unavailable, provide degraded functionality, or fall back to software implementations with reduced performance.

Unsupported APIs#

The following APIs from libcudacxx are not supported in libhipcxx:

Group

API

Description

Synchronization Library

<cuda/std/latch>

Single-phase asynchronous thread-coordination mechanism.

Synchronization Library

<cuda/std/barrier>

Multi-phase asynchronous thread-coordination mechanism.

Synchronization Library

<cuda/std/semaphore>

Primitives for constraining concurrent access.

Extended Synchronization Library

<cuda/latch>

System-wide cuda::std::latch single-phase asynchronous thread coordination mechanism.

Extended Synchronization Library

<cuda/barrier>

System-wide cuda::std::barrier multi-phase asynchronous thread coordination mechanism.

Extended Synchronization Library

<cuda/semaphore>

System-wide primitives for constraining concurrent access.

Extended Synchronization Library

<cuda/pipeline>

Coordination mechanisms to sequence asynchronous operations.

Extended Memory Access Properties Library

<cuda/annotated_ptr>

Memory access properties for pointers.

Extended Memory Access Properties Library

<cuda/discard_memory>

Discards modified cache lines without writing them back. The header is available, but cuda::discard_memory has no effect on AMD GPUs.

Extended Warp Library

<cuda/warp>

Warp shuffle, match and lane-mask functions. The header is available, but cuda::device::warp_shuffle_*, cuda::device::warp_match_all and cuda::device::lane_mask are only compiled for NVIDIA PTX targets.

Extended Work Stealing Library

<cuda/work_stealing>

cuda::for_each_canceled_block for cancelling and stealing thread blocks. On AMD GPUs it invokes the function once for the current block and does not cancel other blocks.

Device-Level APIs

cuda::device::*

Hardware-specific device functions including warp shuffles, barrier operations with transaction counts, and async memory operations. These require NVIDIA PTX instructions and SM-specific hardware features not available on AMD GPUs.

Extended TMA Library

<cuda/tma>

cuda::make_tma_descriptor creates descriptors for the NVIDIA Tensor Memory Accelerator.

Parallel Algorithms

<cuda/std/execution>

Overloads of the standard algorithms that take an execution policy such as cuda::execution::gpu. These overloads are not declared by default. See Parallel algorithms (experimental) for the experimental opt-in.

PTX API

<cuda/ptx>

The cuda::ptx namespace contains functions that map to NVIDIA PTX instructions.

CUDA Tile

Tile mode

Support for compiling libcudacxx headers in NVIDIA CUDA Tile mode.

Runtime API

<cuda/buffer>

cuda::buffer and the other stream-ordered containers.

Runtime API

cuda::device::current_arch_id, cuda::device::current_arch_traits, cuda::device::current_compute_capability

Queries for the architecture of the current device.

Parallel algorithms (experimental)#

The parallel algorithms in <cuda/std/execution> — cuda::std::reduce, sort, transform, and the other overloads that take an execution policy — are not supported on HIP and are not declared by default. Calling one is a compile error (no matching function).

Upstream implements them on top of NVIDIA CUB. hipCUB does not yet provide several of the primitives CUB 3.x expects, including DeviceTransform::Generate, DeviceTransform::TransformIf, and a two-output DevicePartition. libhipcxx carries a shim that supplies the missing pieces over rocPRIM. That shim is experimental: it is not covered by the same testing as the rest of the library, and its API surface may change.

To opt in, define LIBHIPCXX_ENABLE_EXPERIMENTAL_HIPCUB_SHIM when compiling your own code:

hipcc -DLIBHIPCXX_ENABLE_EXPERIMENTAL_HIPCUB_SHIM ...

When building libhipcxx itself, the equivalent CMake option is -DLIBCUDACXX_ENABLE_EXPERIMENTAL_HIPCUB_SHIM=ON. This option only affects libhipcxx’s own build tree. Code that consumes an installed libhipcxx must define the macro in its own build.

The flag will be removed and the algorithms enabled unconditionally once hipCUB provides the missing primitives.