Enum ThroughputMode 

pub enum ThroughputMode {
    ComputeDirect {
        dtype: ElemType,
    },
    ComputeCmma {
        dtype: ElemType,
        config: ComputeCmmaConfig,
    },
    Memory,
    MemoryRead,
    MemoryWrite,
    MemoryWorkingSet {
        access: MemoryAccess,
        bytes: u64,
    },
    Launch,
}

Represents the mode of a throughput computation.

Variants

ComputeDirect

Compute direct calculation without special hardware acceleration.

Fields

dtype: ElemType

The data type of the computation.

ComputeCmma

Compute cmma calculation with CMMA hardware acceleration.

Fields

dtype: ElemType

The data type of the computation.

config: ComputeCmmaConfig

The configuration of the CMMA operation.

Memory

Memory input reads and output writes — a copy, at the default working set. ops_count counts both directions, so this is total traffic across the memory interface.

MemoryRead

Memory input reads only, no store, at the default working set. The ceiling for a kernel that streams data it does not write back — a weight stream, a reduction, a gather. Such a kernel can legitimately exceed Memory, which is why it needs its own probe rather than a correction factor.

MemoryWrite

Memory output writes only, no read, at the default working set. The ceiling for a kernel that streams stores it never reads back: an RNG fill, a memset, a broadcast. Such a kernel can legitimately exceed Memory, which is why it needs its own probe rather than a correction factor.

MemoryWorkingSet

One point of a memory curve: access over a working set of exactly bytes.

bytes is the traffic one pass moves across the interface, the same currency the measured rate is in — a Copy splits it evenly between two buffers of bytes / 2, a Read moves all of it out of one buffer of bytes. So a kernel that touches N bytes asks about N regardless of which access it resembles.

The probe reads cold: every pass moves to a fresh window of a buffer far larger than the working set, so a small size measures what a kernel that size moves rather than the speed of re-reading something already in cache. A size too small to keep the interface busy therefore reports less than the hardware can do, which is not an error — it is the ceiling for a kernel with that little in flight.

Sweeping this over a range of sizes is MemoryCurve.

Fields

access: MemoryAccess

Which directions of traffic the probe issues.

bytes: u64

The bytes moved in one pass.

Launch

Launch overhead measurement.

Implementations

impl ThroughputMode

pub const fn memory_probe(&self) -> Option<(MemoryAccess, u64)>

The access pattern and working set this mode probes, or None for the modes that don’t measure memory.

The one place the single-size variants are mapped onto their access and size, so nothing downstream has to repeat the equivalence.

Trait Implementations

impl Clone for ThroughputMode

fn clone(&self) -> ThroughputMode

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for ThroughputMode

impl Debug for ThroughputMode

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for ThroughputMode

fn deserialize<__D>( __deserializer: __D, ) -> Result<ThroughputMode, <__D as Deserializer<'de>>::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Eq for ThroughputMode

impl Hash for ThroughputMode

fn hash<__H>(&self, state: &mut __H)

where __H: Hasher,

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for ThroughputMode

fn eq(&self, other: &ThroughputMode) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl Serialize for ThroughputMode

fn serialize<__S>( &self, __serializer: __S, ) -> Result<<__S as Serializer>::Ok, <__S as Serializer>::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for ThroughputMode