Error while building for hipBLAS on Windows 11

[phungthomas]

Hello,

I am trying to build llama.cpp for hipBLAS on gfx1100 on Windows 11, but I get this error:

[...]
D:/Documents/llama.cpp/ggml-cuda.cu:2620:81: error: use of undeclared identifier 'hipHostRegisterReadOnly'
    cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
                                                                                ^
D:/Documents/llama.cpp/./ggml-cuda/common.cuh:78:34: note: expanded from macro 'cudaHostRegisterReadOnly'
#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
                                 ^
13 warnings and 1 error generated when compiling for gfx1100.
[...]

Commit 2bf8d0f seems to have brought this issue, as I am able to successfuly build it on previous commits, but it fails on this one,

Here are the commands that I use to build:

SET CC=clang.exe
SET CXX=clang++.exe
cmake .. -G "Ninja" -DLLAMA_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=gfx1100
cmake --build .

[phungthomas]

Changing ggml-cuda.cu, line 2620 from

cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);

to:

cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable);

makes the build complete successfully, but I have no idea what this does.

[slaren]

The read only flag is necessary for using this with mmap, since the file is mapped read only. And that's the whole point of this function. There is a fix merged in whisper.cpp for older CUDA versions in ggml-org/whisper.cpp#2020, it could also be disabled for HIP in the same way.

[ggerganov]

The fix is now synced here: f77261a