lc_dsl

Define a Kernel1D, Kernel2D, or Kernel3D lambda taking BufferVar or image arguments, use dispatch_id() for indexing, then compile with device.compile(kernel) and launch via stream << shader(args).dispatch(size). Include <luisa/dsl/syntax.h> for core DSL types.

Use the LUISA_STRUCT macro with the struct name and member list, e.g. LUISA_STRUCT(Material, albedo, roughness) {}. The macro body can also define DSL methods, and template structs use LUISA_TEMPLATE_STRUCT with template definition macros.

Yes, it provides warp vote, reduction, prefix scan, and lane exchange intrinsics such as warp_active_sum, warp_active_bit_mask, warp_prefix_sum, and warp_read_lane. Set the warp size with set_warp_size(32u) and query lanes with warp_lane_id().

Cooperative vector operations currently only support the Vulkan backend. The DX backend requires Shader Model 6.8 with experimental features enabled through a DirectXDeviceConfigExt subclass returning true from UseExperimental().

Call buf.atomic(index) to get an atomic accessor, then use fetch_add, fetch_sub, fetch_max, or compare_exchange. Atomics also work on vector components, matrix elements, struct members, and Shared memory arrays.

The $ macros from <luisa/dsl/sugar.h> are shorthand: $float declares Var<float>, $if/$for/$while replace if_ and loop calls, and $buffer declares buffer arguments. Both forms compile identically; sugar reduces boilerplate in kernel code.