ruprim/reduce/routines/
blueprint.rs

1use crate::reduce::{BoundChecks, IdleMode, VectorizationMode};
2
3#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
4pub struct ReduceBlueprint {
5    /// How vectorization was applied.
6    pub vectorization_mode: VectorizationMode,
7    /// The global blueprint for the kernel.
8    pub global: GlobalReduceBlueprint,
9}
10
11#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
12pub enum GlobalReduceBlueprint {
13    Unit(UnitReduceBlueprint),
14    Plane(PlaneReduceBlueprint),
15    Ruda(RudaBlueprint),
16}
17
18#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
19/// A single ruda reduces a full vector.
20pub struct RudaBlueprint {
21    /// When too many rudas are spawned, we should put some to idle.
22    ///
23    /// # Notes
24    ///
25    /// This only happens when we hit the hardware limit in spawning rudas on a single axis.
26    pub ruda_idle: IdleMode,
27    /// There are too many units in a ruda causing out-of-bound.
28    ///
29    /// # Notes
30    ///
31    /// There are never too many rudas spawned.
32    pub bound_checks: BoundChecks,
33    /// The number of accumulators in shared memory.
34    pub num_shared_accumulators: usize,
35    /// Whether we use plane instructions to merge accumulators.
36    pub use_planes: bool,
37}
38
39#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
40pub enum PlaneMergeStrategy {
41    ///  All units in a plane work independently during the reduction
42    ///  but merge their accumulators at the end
43    Lazy,
44    ///  There is a plane reduction at each iteration
45    Eager,
46}
47
48/// A single plane reduces a full vector.
49#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
50pub struct PlaneReduceBlueprint {
51    /// Too many planes are spawned, we should put some to idle.
52    pub plane_idle: IdleMode,
53    /// There are too many units in a plane causing out-of-bound.
54    pub bound_checks: BoundChecks,
55    /// Whether we recombine accumulators at each iteration or at the end only
56    pub plane_merge_strategy: PlaneMergeStrategy,
57    /// If true, we ceil the used ruda_dim x to runtime plane_dim
58    pub plane_dim_ceil: bool,
59}
60
61/// A single unit reduces a full vector.
62#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
63pub struct UnitReduceBlueprint {
64    // Too many units are spawned, we should put some to idle.
65    pub unit_idle: IdleMode,
66}
ruprim/reduce/routines/blueprint.rs

ruprim/reduce/routines/
blueprint.rs
