Introduction

Overview

hipCUB is a thin wrapper library on top of rocPRIM or CUB. It enables developers to port project using CUB library to the HIP layer and to run them on AMD hardware. In the ROCm environment, hipCUB uses rocPRIM library as the backend, however, on CUDA platforms it uses CUB instead.

  • When using hipCUB you should only include <hipcub/hipcub.hpp> header.

  • When rocPRIM is used as backend HIPCUB_ROCPRIM_API is defined.

  • When CUB is used as backend HIPCUB_CUB_API is defined.

  • Backends are automaticaly selected based on platform detected by HIP layer (__HIP_PLATFORM_HCC__, __HIP_PLATFORM_NVCC__).

rocPRIM backend

hipCUB with rocPRIM backend may not support all function and features CUB has because of the differences between ROCm (HIP) platform and CUDA platform.

Not-supported features and differences:

  • Functions, classes and macros which are not in the public API or not documented are not supported.

  • Device-wide primitives can’t be called from kernels (dynamic parallelism is not supported in HIP on ROCm).

  • DeviceSpmv is not supported.

  • Fancy iterators: CacheModifiedInputIterator, CacheModifiedOutputIterator, and TexRefInputIterator are not supported.

  • Thread I/O:

    • CacheLoadModifier, CacheStoreModifier cache modifiers are not supported.

    • ThreadLoad, ThreadStore functions are not supported.

  • Storage management and debug functions:

    • Debug, PtxVersion, SmVersion functions and CubDebug, CubDebugExit, _CubLog macros are not supported.

  • Intrinsics:

    • ThreadExit, ThreadTrap - not supported.

    • Warp thread masks (when used) are 64-bit unsigned integers.

    • member_mask input argument is ignored in WARP_* functions.

    • Arguments first_lane, last_lane, and member_mask are ignored in Shuffle* functions.

  • Utilities:

    • SwizzleScanOp, ReduceBySegmentOp, ReduceByKeyOp, CastOp - not supported.