[ROCm] Add CINN HIP BF16 codegen support

[austin1997]

PR Category

Custom Device

PR Types

Bug fixes

Description

Add BF16 type support to CINN HIP device code generation.

This PR:

• Includes bfloat16.h in the HIP device source header and exports CINN bfloat16 vector aliases.
• Adds a CINN_WITH_HIP branch in cinn/common/bfloat16.h that uses ROCm hip_bfloat16 for device-side float/BF16 conversion when HIP_VERSION >= 60100000.
• Adds a ROCm-guarded header coverage check in the CINN compilation task test source.

Verification:

• cmake .
• env TARGET=SKYLAKEX ninja -j 160 paddle_python
• HIPCC compile test for CINN_WITH_HIP bfloat16 construction, conversion, and to_hip_bfloat16 on gfx942
• prek run --files paddle/cinn/backends/hip/codegen_hip_dev.cc paddle/cinn/common/bfloat16.h test/cpp/pir/cinn/compilation_task_test.cc
• git diff --check

Note: the current local build configuration does not generate test_compilation_task in CTest (ctest -N -R test_compilation_task reports 0 tests), so the added gtest is covered when CINN tests are enabled.

是否引起精度变化

否