Language reference

Every builtin in enceladus.language, imported as tl, and the host API in enceladus. Tile shapes must be compile-time powers of two. Most math and shape functions are also tile methods, for example x.exp(), x.sum(axis=0), x.to(tl.float16), and x.T.

Builtins

Data types

The following table lists the element types. Pass them as tl.float16 in a kernel, or as the string "float16" to the host allocators.

Kind	Types	Short names
Boolean	`tl.int1`	`i1`, `bool`
Signed integer	`tl.int8`, `tl.int16`, `tl.int32`, `tl.int64`	`i8`, `i16`, `i32`, `i64`
Unsigned integer	`tl.uint8`, `tl.uint16`, `tl.uint32`, `tl.uint64`	`u8`, `u16`, `u32`, `u64`
Floating point	`tl.float16`, `tl.bfloat16`, `tl.float32`	`fp16`, `bf16`, `fp32`

Environment variables

Variable	Effect
`ENCELADUS_INTERPRET=1`	Runs kernels in the NumPy interpreter on the CPU.
`ENCELADUS_DEBUG=1`	Compiles `tl.device_assert` checks. It's part of the specialization, so changing it recompiles.
`ENCELADUS_DUMP=1`	Writes each compiled kernel's IR, MSL, and `meta.json` to its cache entry and prints the path.
`ENCELADUS_CACHE_DIR`	Moves the disk cache from `~/.cache/enceladus`.
`ENCELADUS_ALWAYS_COMPILE=1`	Ignores the disk cache and saved autotuning results.
`ENCELADUS_OVERRIDE_DIR`	Loads `DIR/KERNEL_NAME.metal` instead of the generated MSL.
`ENCELADUS_VERIFY=1`	Verifies the IR after every compiler pass, including for interpreted launches.
`ENCELADUS_PRINT_AUTOTUNING=1`	Prints each autotuning winner as a pasteable `enceladus.Config(...)`.
`ENCELADUS_FLUSH_EVERY`	Sets the number of launches per command buffer. The default is 64.
`MTL_CAPTURE_ENABLED=1`	A Metal variable that `enceladus.capture` needs at process start.