Compiles LuisaCompute AST kernels to Vulkan SPIR-V through LLVM's SPIR-V target.
npx skills add https://github.com/LuisaGroup/LuisaCompute --skill llvm-spirv
Please help me install this Agent Skill. Skill: llvm-spirv Source: https://github.com/LuisaGroup/LuisaCompute/tree/main/.agents/skills/llvm-spirv Command: npx skills add https://github.com/LuisaGroup/LuisaCompute --skill llvm-spirv
It provides an experimental alternative Vulkan codegen path that lowers LuisaCompute AST functions to LLVM IR and emits SPIR-V modules, while enforcing a fail-closed runtime-interface boundary so unsupported kernels are rejected instead of silently miscompiled.
spirv64-unknown-vulkan1.2 module via the LLVM-free public facade lc::llvm_codegen::compile_spirv.validate_llvm_vulkan_resource_model rejects kernel arguments, textures, bindless arrays, acceleration structures, and shader printing before lowering, so only no-argument, no-print compute kernels compile.luisa-compute-spirv-llvm) and XMake (lc-spirv-llvm) configuration with component-aware LLVM discovery, ABI checks, and private LLVM headers.Compile a no-argument compute kernel through the public facade and validate the returned SPIR-V module with SPIRV-Tools under SPV_ENV_VULKAN_1_2.
High-intent search queries and answers about installing and using this skill.
Call lc::llvm_codegen::compile_spirv(kernel, option) through the public header spirv_llvm/spirv_llvm.h. The facade runs target registration, AST lowering, aggregate legalization, post-processing, and Vulkan validation, returning SPIR-V words and properties.
They are mutually exclusive Vulkan codegen paths selected at build time. The LLVM path lowers AST through LLVM IR, embeds constants in the module, and uses the LLVM_SPIRV artifact dialect, while the native XIR path has its own optimizer and exact per-argument role plan.
Only no-argument, no-print compute kernels whose operations are implemented by the visitor. The resource preflight rejects all kernel arguments, textures, bindless arrays, acceleration structures, custom resources, and shader printing before lowering.
In CMake set LUISA_COMPUTE_ENABLE_VK_AST_LLVM_SPIRV=ON and LUISA_COMPUTE_ENABLE_VK_XIR_SPIRV=OFF. In XMake set lc_vk_backend_use_ast_llvm_spirv=true with lc_llvm_path pointing to an LLVM prefix or llvm-config.
The property preflight validate_llvm_vulkan_resource_model intentionally rejects every kernel argument because no real descriptor ABI exists yet. Provisional global-variable lowering exists, but the preflight blocks those kernels until the Vulkan binding shape is validated.
No. The facade must stay LLVM-header-free so consumers do not inherit LLVM include directories or preprocessor definitions. Keep LLVM includes private and fix link failures through proper component linkage instead.
hw-native-sysBuild and validate the PTOAS compiler toolchain from source in WSL.
alibabaOptimizes MNN Vulkan backend kernels and integrates new Vulkan features on mobile GPUs.
bitwize-music-studioExecute pytest suites across 14 categories to validate Claude Code plugin integrity.
nebinfraAutomate end-to-end live testing of NebCore AI platform across AI server, backend, and agent runtimes.