Struct ToleranceTable 

pub struct ToleranceTable {
    pub vector_add_f32: Tolerance,
    pub vector_mul_f32: Tolerance,
    pub vector_fma_f32: Tolerance,
    pub matmul_f32: Tolerance,
    pub conv2d_f32: Tolerance,
    pub f16_ulp_multiplier: u32,
}

Per-blueprint floating-point tolerance for CPU↔GPU bit-identity checks.

fp32: 1 ULP allowed for element-wise ops; 2 ULP for reductions (matmul). fp16: 4 ULP allowed (smaller mantissa, larger drift). Integer: bit-identical required (tolerance == 0).

Fields

vector_add_f32: Tolerance

Vector add tolerance keyed by dtype.

vector_mul_f32: Tolerance

Vector mul tolerance.

vector_fma_f32: Tolerance

FMA tolerance.

matmul_f32: Tolerance

Matmul tolerance (looser — reduction order is impl-defined).

conv2d_f32: Tolerance

Conv2d tolerance (matmul-shaped budget).

f16_ulp_multiplier: u32

f16 multiplier — every f32 budget is scaled by this for f16.

Implementations

impl ToleranceTable

pub const fn strict() -> Self

Strict policy: zero tolerance everywhere. Used by the CI gate that asserts “the JIT produced identical bytes” — for ops the oracle pins as bit-identical (today: integer kernels only).

pub const fn default_table() -> Self

Default per-blueprint policy. Values mirror the table in docs/DIFFERENTIAL-ORACLE.md#tolerance-table-v037.

pub fn for_blueprint(&self, kind: BlueprintKind, dtype: Dtype) -> Tolerance

Look up the tolerance for the given blueprint kind + dtype. Integer kernels are always strict; f16 multiplies the f32 ULP budget by Self::f16_ulp_multiplier.

Trait Implementations

impl Clone for ToleranceTable

fn clone(&self) -> ToleranceTable

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ToleranceTable

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ToleranceTable

fn default() -> Self

Returns the “default value” for a type.