Add RISC-V QEMU test coverage

[carlosqwqqwq]

Why

FP16 already has a portable scalar implementation for its conversion routines, and the existing architecture-specific fast paths are optional optimizations rather than hard requirements.

For this project, the missing part is official riscv64 verification coverage:

• the current code already falls back to a conservative non-native path when F16C, AArch64 FP16, and related native conversion options are unavailable;
• that conservative path can be built for riscv64 today, but upstream CI did not exercise it;
• without a real cross-compiled riscv64 test run, the portability claim stays weaker than it needs to be.

This patch keeps the conversion logic unchanged and adds a reproducible riscv64 Linux + QEMU test path to the existing CMake workflow.

What changed

• Updated .github/workflows/cmake.yml:
  • add a cmake-linux-riscv64 job on ubuntu-24.04;
  • install g++-riscv64-linux-gnu, qemu-user-static, and ninja-build;
  • configure CMake for riscv64 cross-compilation with CMAKE_CROSSCOMPILING_EMULATOR;
  • build and run the full unit-test suite under qemu-riscv64-static;
  • disable benchmarks in the riscv64 job to keep the job focused on correctness and runtime manageable.
• Updated README.md:
  • document that GitHub Actions CI also cross-compiles and runs the unit tests for riscv64 Linux under QEMU.

Verification

• Verified native configure/build/test on the Windows host:
  • cmake -S . -B build-native -G Ninja -DCMAKE_BUILD_TYPE=Release -DFP16_BUILD_BENCHMARKS=OFF -DFP16_BUILD_COMPARATIVE_BENCHMARKS=OFF
  • cmake --build build-native --parallel 4
  • ctest --test-dir build-native --output-on-failure
  • result: 7/7 tests passed.
• Verified the conservative RISC-V macro path:
  • forced __riscv=1 / __riscv_xlen=64 compile probe against include/fp16/macros.h and include/fp16/fp16.h;
  • confirmed FP16_USE_NATIVE_CONVERSION == 0, FP16_USE_FLOAT16_TYPE == 0, and FP16_USE_FP16_TYPE == 0 on the conservative RISC-V path.
• Verified real riscv64 Linux cross-compilation and QEMU execution in Docker (ubuntu:24.04):
  • configured CMake with riscv64-linux-gnu-gcc, riscv64-linux-gnu-g++, CMAKE_SYSTEM_PROCESSOR=riscv64, and CMAKE_CROSSCOMPILING_EMULATOR=qemu-riscv64-static;-L;/usr/riscv64-linux-gnu;
  • built the test suite successfully;
  • ran the full ctest suite under QEMU with no failures.
• Verified the completed QEMU run from CTest logs:
  • Start testing: Jun 18 13:43 UTC
  • End testing: Jun 18 14:00 UTC
  • PassedCount = 7
  • FailedCount = 0
  • LastTest.log confirms execution through qemu-riscv64-static.
• Verified the cross build configuration:
  • build-riscv/CMakeCache.txt records CMAKE_C_COMPILER=/usr/bin/riscv64-linux-gnu-gcc
  • CMAKE_CXX_COMPILER=/usr/bin/riscv64-linux-gnu-g++
  • CMAKE_SYSTEM_PROCESSOR=riscv64
  • CMAKE_CROSSCOMPILING_EMULATOR=qemu-riscv64-static;-L;/usr/riscv64-linux-gnu
• Verified the workflow file parses as valid YAML.

Notes

• This patch does not change the FP16 conversion algorithms or add a RISC-V-specific optimized backend.
• The goal is to make the already-portable scalar path officially verified for riscv64, not to claim new architecture-specific acceleration.