vyre 0.4.0

GPU compute intermediate representation with a standard operation library
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63

use crate::ir::{Expr, Program};
use crate::ops::primitive;
use crate::ops::{AlgebraicLaw, OpSpec, F32_F32_INPUTS, F32_OUTPUTS};

// Element-wise multiplication of two f32 buffers.


// IEEE-754 f32 multiplication is NOT bitwise-associative under
// rounding and `Absorbing { element: 0 }` fails on `0 * NaN = NaN` and
// `0 * Inf = NaN` (neither returns 0). `Identity { element: 1 }` holds
// bit-exactly (IEEE guarantees `1 * x = x` for every finite, signed
// zero, infinity, and NaN payload). Per audit
// `primitive-laws-kimi-AUDIT`, F32Mul satisfies commutativity and
// right-identity-by-one only.
pub const LAWS: &[AlgebraicLaw] = &[
    AlgebraicLaw::Commutative,
    AlgebraicLaw::Identity { element: 1.0f32.to_bits() },
];

/// f32 element-wise multiplication operation.
#[derive(Debug, Clone, Copy, Default)]
pub struct F32Mul;

impl F32Mul {
    /// Declarative operation specification.
    pub const SPEC: OpSpec = OpSpec::composition_inlinable(
        "primitive.float.f32_mul",
        F32_F32_INPUTS,
        F32_OUTPUTS,
        LAWS,
        Self::program,
    );

    /// Build the canonical IR program.
    ///
    /// # Examples
    ///
    /// ```
    /// use vyre::ir::Expr;
    /// use vyre::ops::primitive::float::f32_mul::F32Mul;
    ///
    /// let program = F32Mul::program();
    /// assert!(!program.entry().is_empty());
    /// ```
    #[must_use]
    pub fn program() -> Program {
        primitive::binary_f32_program(Expr::mul)
    }
}

// Backend-specific lowering tripwires for `primitive.float.f32_mul`.

// WGSL lowering marker for `primitive.float.f32_mul`.
//
// Not a stub: this is a zero-overhead Category A marker. `F32Mul::program`
// builds concrete IR through `core/src/ops/primitive/binary_f32_program.rs`;
// `core/src/lower/wgsl/expr.rs::emit_binop` emits WGSL.
// `core/tests/conformance.rs::conformance_all_primitives` verifies
// lowered GPU bytes are bit-exact against the conform CPU reference.
//
// ```wgsl
// _vyre_store_out(idx, (_vyre_load_a(idx) * _vyre_load_b(idx)));
// ```