Skip to main content

LowerOp

cubecl_cpp::shared::lowering

Trait LowerOp

pub trait LowerOp<T = Shared>: Op {
    // Required method
    fn lower(&self, scope: &Scope) -> Vec<Value>;

    // Provided methods
    fn verify(_op: &dyn Op, _ctx: &Context) -> Result<()>
       where Self: Sized { ... }
    fn should_lower(&self, _ctx: &Context) -> bool { ... }
}

Required Methods§

fn lower(&self, scope: &Scope) -> Vec<Value>

Provided Methods§

fn verify(_op: &dyn Op, _ctx: &Context) -> Result<()>
where Self: Sized,

fn should_lower(&self, _ctx: &Context) -> bool

Trait Implementations§

impl<T> OpInterfaceMarker for dyn LowerOp<T>

Dyn Compatibility§

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementations on Foreign Types§

impl LowerOp for AtomicFSubOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for AtomicISubOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for AtomicStoreOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for AtomicUniformLoadOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for CastOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for DeclareVariableOp

fn should_lower(&self, _ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for DegreesOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for Dp4aOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for ErfOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for ExclusiveFProdOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for ExclusiveFSumOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for ExclusiveIProdOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for ExclusiveISumOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for FDotOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for FMaxOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for FMinOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for FProdOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for FSumOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for FSumOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for FindFirstSetOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for IProdOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for ISumOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for ISumOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for InclusiveFProdOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for InclusiveFSumOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for InclusiveIProdOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for InclusiveISumOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for Log1pOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for MagnitudeOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for NormalizeOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for PowiOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for RadiansOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for RecipOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for SDotOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for SMaxOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for SMinOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for TrailingZerosBitsOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for UDotOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for UMaxOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for UMinOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp for UniformLoadOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for CastOp

special cast function for recursive conversion in the case of minifloat to minifloat conversion

Needs to jump through a lot of hoops to deal with CUDA nonsense. The overview of available conversions is as follows:

From	To	Extra args
f16/bf16/f32/f64	e4m3/e5m2	Interpretation, saturation
f16/bf16/f32/f64	e3m2/e2m3/e2m1	Interpretation, rounding
bf16/f32/f64	e8m0	saturation, rounding
e4m3/e5m2/e3m2/e2m3/e2m1	f16	Interpretation,
e8m0	bf16

When the input and output don’t match these options, we need to do a two-step conversion. When the input is a minifloat we always need to cast out to f16/bf16, and then convert to the actual out type if it differs. Trying to cast ints also requires an extra conversion, and so does f16 to e8m0 (though it’s not recommended to do that anyways, you should be using e5m2 for that since you don’t have 8 bits of exponent in f16).

See also: https://docs.nvidia.com/cuda/cuda-math-api/cuda_math_api/group__CUDA__MATH__FP8__MISC.html https://docs.nvidia.com/cuda/cuda-math-api/cuda_math_api/group__CUDA__MATH__FP6__MISC.html https://docs.nvidia.com/cuda/cuda-math-api/cuda_math_api/group__CUDA__MATH__FP4__MISC.html

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for ColIndexOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for CommitCopyAsyncOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for CopyAsyncOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for Dp4aOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for ElectOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for InlineAsmOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for LdMatrixOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for MmaManualOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for MmaManualScaledOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for RowIndexOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for SaturatingSAddOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for SaturatingSSubOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for StMatrixOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Cuda> for TmaLoadIm2colOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Hip> for BallotOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Hip> for ColIndexOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Hip> for ElectOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Hip> for FMaxOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Hip> for FMinOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Hip> for RowIndexOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Hip> for TanhOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Metal> for BallotOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Metal> for Expm1Op

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Metal> for SMulHiOp

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Metal> for TanhOp

fn should_lower(&self, ctx: &Context) -> bool

fn lower(&self, scope: &Scope) -> Vec<Value>

impl LowerOp<Metal> for UMulHiOp

fn lower(&self, scope: &Scope) -> Vec<Value>

Implementors§