Compile bug: RISC-V: vector still be enabled even if GGML_RVV=OFF #16593
Description
Git commit
Operating systems
Linux
GGML backends
HIP
Problem description & steps to reproduce
Compiler: Clang 20.0.0git
OS: Debian Trixie
I am using llvm-project from AMD ROCm Org to build llama.cpp on a riscv64 machine which does not support RISC-V vector extension.
The code mentioned by the log are guarded by __riscv_v_intrinsic definition. But __riscv_v_intrinsic should not be used to test if compilation host support RVV since llvm 20.1.0-rc1.
Is it better to use __riscv_vector to guard these code?
First Bad Commit
Compile command
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build \
-DCMAKE_C_COMPILER=/opt/rocm-7.0.0/bin/clang -DCMAKE_CXX_COMPILER=/opt/rocm-7.0.0/bin/clang++ \
-DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1100 \
-DCMAKE_VERBOSE_MAKEFILE=TRUE \
-DGGML_RVV=OFF -DGGML_RV_ZFH=OFF -DGGML_RV_ZICBOP=OFF \
-DCMAKE_BUILD_TYPE=ReleaseCPU Info
$ cat /proc/cpuinfo | tail | grep 'hart isa'
hart isa : rv64imafdch_zicntr_zicsr_zifencei_zihpm_zaamo_zalrsc_zca_zcdRelevant log output
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:365:31: error: RISC-V type '__rvv_float32m8_t' requires the 'zve32f' extension
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:365:31: error: RISC-V type '__rvv_float32m8_t' requires the 'zve32f' extension
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:365:55: error: RISC-V type 'vfloat32m8_t' (aka '__rvv_float32m8_t') requires the 'zve32f' extension
365 | vfloat32m8_t ny = __riscv_vfmadd_vf_f32m8(ax, v, ay, avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:365:62: error: RISC-V type 'vfloat32m8_t' (aka '__rvv_float32m8_t') requires the 'zve32f' extension
365 | vfloat32m8_t ny = __riscv_vfmadd_vf_f32m8(ax, v, ay, avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:366:13: error: RISC-V type '__rvv_float32m8_t' requires the 'zve32f' extension
366 | __riscv_vse32_v_f32m8(&y[i], ny, avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:366:42: error: RISC-V type 'vfloat32m8_t' (aka '__rvv_float32m8_t') requires the 'zve32f' extension
366 | __riscv_vse32_v_f32m8(&y[i], ny, avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:537:26: error: RISC-V type 'vfloat32m8_t' (aka '__rvv_float32m8_t') requires the 'zve32f' extension
537 | vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:537:31: error: RISC-V type '__rvv_float32m8_t' requires the 'zve32f' extension
537 | vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:539:30: error: RISC-V type 'vfloat32m8_t' (aka '__rvv_float32m8_t') requires the 'zve32f' extension
539 | vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[k][i], avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:539:35: error: RISC-V type '__rvv_float32m8_t' requires the 'zve32f' extension
539 | vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[k][i], avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:540:17: error: RISC-V type 'vfloat32m8_t' (aka '__rvv_float32m8_t') requires the 'zve32f' extension
540 | ay = __riscv_vfmadd_vf_f32m8(ax, v[k][0], ay, avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:540:22: error: RISC-V type '__rvv_float32m8_t' requires the 'zve32f' extension
540 | ay = __riscv_vfmadd_vf_f32m8(ax, v[k][0], ay, avl);
| ^
/home/debian/Sakura286/llama/llama.cpp/ggml/src/ggml-cpu/vec.h:540:22: error: RISC-V type '__rvv_float32m8_t' requires the 'zve32f' extension
fatal error: too many errors emitted, stopping now [-ferror-limit=]