tiny.gpu.BackendMathTier

1/// A caller picks the math tier, the arithmetic contract compiled code follows, before compiling2/// for a device. Two choices exist: `exact`, the default, and `tf32_tensor`, which uses the3/// reduced-precision TF32 format on NVIDIA tensor cores. The option `tf32_tensor` is accepted only4/// for a CUDA backend, one device runtime behind the common interface, emitting PTX on a device5/// with tensor cores. Compiling fails with an error when a kernel cannot be compiled under6/// `tf32_tensor`, and the code never falls back to another tier on its own. The chosen tier is7/// hashed into the identity of the compiled artifact, so code compiled under one tier remains8/// distinct from code compiled under the other. A tier chooses arithmetic precision alone, and it9/// gives no license for algebraic rewrites that change which values are NaN, infinities or the sign10/// of zero.11pub const BackendMathTier = enum(u8) {12    exact = 0,13    tf32_tensor = 1,1415    pub fn parse(text: []const u8) ?BackendMathTier {16        inline for (17            @typeInfo(BackendMathTier).@"enum".field_names,18            @typeInfo(BackendMathTier).@"enum".field_values,19        ) |field_name, field_name_value| {20            const field = .{ .name = field_name, .value = field_name_value };21            const value: BackendMathTier = @fromBackingInt(@intCast(field.value));22            if (std.mem.eql(u8, text, @tagName(value))) return value;23        }24        return null;25    }26};

1pub const BackendMathTier = contract.BackendMathTier;