ci: add Linux cross-compile build

[bandoti]

This change introduces a cross-compile build targeting Riscv on Linux. The goal is to reduce regression issues associated with cross-compiling and to serve as an example how to cross-compile using Ubuntu. In the future we can update this to store artifacts if there is a larger demand to run on Riscv (or other) hardware.

[bandoti]

@jeffbolznv I am having an issue with the cross-compile build missing an extension. Here I arbitrarily selected Riscv for the cross-compile, and although I can pull libvulkan-dev:riscv64, I am wondering whether the version in the LunarG SDK is farther along and I cannot build without newer features.

/home/runner/work/llama.cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp:940:9: error: ‘PipelineRobustnessCreateInfoEXT’ is not a member of ‘vk’; did you mean ‘PipelineColorWriteCreateInfoEXT’?
  940 |     vk::PipelineRobustnessCreateInfoEXT rci;
      |         ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
      |         PipelineColorWriteCreateInfoEXT
/home/runner/work/llama.cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp:943:9: error: ‘rci’ was not declared in this scope
  943 |         rci.storageBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled;
      |         ^~~
/home/runner/work/llama.cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp:943:34: error: ‘vk::PipelineRobustnessBufferBehaviorEXT’ has not been declared
  943 |         rci.storageBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled;
      |                                  ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

[jeffbolznv]

Pipeline robustness is a pretty old extension at this point - changelog says "July 14, 2022 Vulkan 1.3.221". I had recently seen 1.3.261.1 SDK used in one of the CI jobs, so maybe there's an even older version used in another job?

We can potentially change the code to make it not rely on this, but it would be better to get all builds on a more recent SDK.

[bandoti]

Looks like updating to "ubuntu-latest" may have fixed that issue. I'm seeing a different one, which might indicate an issue with the generated shaders.

[ 20%] Building CXX object ggml/src/ggml-vulkan/CMakeFiles/ggml-vulkan.dir/ggml-vulkan-shaders.cpp.o
/home/runner/work/llama.cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp: In function ‘void ggml_vk_load_shaders(vk_device&)’:
/home/runner/work/llama.cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp:1876:55: error: ‘matmul_f32_f32_coopmat_len’ was not declared in this scope; did you mean ‘matmul_f32_f32_fp32_len’?
 1876 |         CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
      |                                                       ^~~~~~~~~~~~~~
/home/runner/work/llama.cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp:1855:95: note: in definition of macro ‘CREATE_MM’
 1855 |             ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true);   \
      |                                                                                               ^~~~~~
/home/runner/work/llama.cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp:1876:55: error: ‘matmul_f32_f32_coopmat_data’ was not declared in this scope; did you mean ‘matmul_f32_f32_fp32_data’?
 1876 |         CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
      |                                                       ^~~~~~~~~~~~~~
/home/runner/work/llama.cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp:1855:129: note: in definition of macro ‘CREATE_MM’
 1855 |             ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true);   \
      |                                                                                                                                 ^~~~~~

[jeffbolznv]

This looks similar to issues like #11695. I suspect these are due to using a different glslc in the two environments (i.e. compilation actually happens with an older glslc than the feature detection does). Maybe there's some logging you can add to the cmake to see if that's happening?