Data Types

All scalar types in Enigma with their Metal and numpy equivalents.

Enigma defines a fixed set of scalar types that map directly to Metal Shading Language types. Use them as kernel parameter type annotations and as arguments to casting functions.

Scalar types

Enigma type	Metal type	Bits	Python/NumPy
enigma.f32	float	32	np.float32
enigma.f16	half	16	np.float16
enigma.bf16	bfloat	16	ml_dtypes.bfloat16
enigma.i32	int	32	np.int32
enigma.u32	uint	32	np.uint32
enigma.i64	long	64	np.int64
enigma.u64	ulong	64	np.uint64
enigma.i16	short	16	np.int16
enigma.u16	ushort	16	np.uint16
enigma.i8	char	8	np.int8
enigma.u8	uchar	8	np.uint8
enigma.b1	bool	1	—

Using types as kernel parameters

Type annotations on kernel parameters declare the buffer element type:

@enigma.kernel
def mixed(A: enigma.f16, B: enigma.i32, C: enigma.f32):
    tid = enigma.thread_position_in_grid
    val = enigma.metal_cast(A[tid], enigma.f32)
    idx = B[tid]
    C[idx] = val

Type casting

enigma.metal_cast(value, target_type)

Standard type conversion, equivalent to C-style casting in Metal:

x_f32 = enigma.metal_cast(x_f16, enigma.f32)   # half → float
x_i32 = enigma.metal_cast(x_f32, enigma.i32)   # float → int (truncate)

enigma.as_type(value, target_type)

Bitwise reinterpret (no value conversion, same bit pattern):

bits = enigma.as_type(x_f32, enigma.u32)        # float bits as uint32

Integer literals

Integer literals in kernel bodies are automatically promoted to u32:

idx = tid * 4 + 1   # 4 and 1 become u32 IRValues

Float literals become f32:

val = A[tid] * 2.0  # 2.0 is f32

Scalar parameters

enigma.Scalar(dtype) declares a per-dispatch constant. It is lowered as a 1-element device buffer and auto-loaded at kernel entry:

@enigma.kernel
def add_scalar(A: enigma.f32, bias: enigma.Scalar(enigma.f32), C: enigma.f32):
    tid = enigma.thread_position_in_grid
    C[tid] = A[tid] + bias

Pass the scalar value as a 1-element numpy array at dispatch time:

rt.execute(compiled, inputs=[a, np.array([0.5], dtype=np.float32)], ...)

Vector types

Enigma supports vector construction for geometric operations. Vector types are created from scalars:

v2 = enigma.make_float2(x, y)
v3 = enigma.make_float3(x, y, z)
v4 = enigma.make_float4(x, y, z, w)

Access components via .x, .y, .z, .w, or by lane index:

px = v3.x
lane1 = enigma.vec_extract(v3, 1)  # equivalent to .y

See Operations for geometric and vector math functions.