ROCm/hip
Hip
HIP is a C++ runtime API and kernel language for writing parallel GPU code that runs on AMD GPUs and can also target NVIDIA GPUs. It is part of the AMD ROCm GPU compute platform and lets host code and GPU kernels live together in ordinary C++ source files, which a Clang-based compiler splits into CPU and GPU portions at build time. Kernels are launched with an inline grid-and-block syntax, and the runtime covers launching kernels, managing device memory and coordinating work between CPU and GPU.
What sets HIP apart is its close mirroring of the CUDA runtime API. Existing CUDA applications can be moved to AMD hardware largely through mechanical renaming, without a full rewrite. The same codebase can build for either AMD or NVIDIA hardware, with runtime calls forwarded through a thin dispatch layer to the native AMD runtime or an NVIDIA backend.
The programming model organizes work into a hierarchy of thread blocks within a grid, with per-block shared memory and barriers for cooperative computation. Kernel launches and memory copies are queued onto ordered streams that run independently of the CPU, and events are used to synchronize them. Many device functions and intrinsics are supplied as inline headers that the compiler expands directly into GPU instructions. Typical uses include heterogeneous CPU-GPU applications and compute-heavy scientific workloads.
