Lowers XIR kernels to validated SPIR-V 1.5 for the Vulkan backend.
npx skills add https://github.com/LuisaGroup/LuisaCompute --skill spirv-codegen
Please help me install this Agent Skill. Skill: spirv_codegen Source: https://github.com/LuisaGroup/LuisaCompute/tree/main/.agents/skills/spirv_codegen Command: npx skills add https://github.com/LuisaGroup/LuisaCompute --skill spirv-codegen
It guides work on LuisaCompute's native XIR-to-SPIR-V code generator, where incorrect legalization, descriptor binding, or control-flow emission produces invalid Vulkan shaders that fail validation or misbehave at runtime.
compile_spirv, compile_spirv_xir), the mandatory AST-to-XIR legalization pipeline, and the fail-closed dialect validation boundary.ControlFlowPlan, native OpPhi emission, and rules for loop headers, merges, and forwarding blocks.SPV_ENV_VULKAN_1_2 validation.Ask the AI to add support for a new XIR opcode in the native SPIR-V codegen following the dialect boundary and validation rules in this skill.
High-intent search queries and answers about installing and using this skill.
Classify the opcode in `validate_spirv_xir_codegen_dialect` as supported, semantic no-op, or unsupported, then validate operand counts, types, and storage classes. Add a focused diagnostic, update the opcode-matrix test, and add an exact codegen test if accepted.
The native path in `spirv_codegen/` lowers XIR directly with glslang's `spv::Builder`, while `spirv_llvm/` implements a separate AST-to-LLVM-to-SPIR-V pipeline. This skill covers only the native path enabled by `LUISA_COMPUTE_ENABLE_VK_XIR_SPIRV`.
The pre-optimization module must already be valid; the optimizer is never a repair step. Check dialect handoff violations such as mismatched load/store types, invalid texture texel widths, or Phi placement rules before blaming SPIRV-Tools.
No. Callable ABI failures are solved by fixed-point specialization of buffer, bindless, and resource arguments at call sites, not by enabling `VariablePointers` globally. Only safe opaque/resource modes remain as callable parameters.
Emission calls `_require_target_feature` at each semantic site, then capability-owned bits are reconciled from the final validated binary while emission-owned bits stay provisional. The persisted mask is checked exactly against the enabled logical-device features on load.
Never in this backend. The runtime exposes no specialization-constant ABI, so ordinary XIR arithmetic emits ordinary SPIR-V instructions, and constant folding belongs in the XIR or SPIRV-Tools optimization layers.
xchang1121Diagnose Triton Ascend kernel issues with a structured debugging checklist.
alibabaOptimizes MNN Vulkan backend kernels and integrates new Vulkan features on mobile GPUs.
haolangeParse HLSL, GLSL, and SPIR-V to detect suspicious shader fingerprints.
simstudioaiDetects and fixes unnecessary useState, derived state, and server-state anti-patterns in React code.