cubek_convolution/components/global/read/strategy/
async_full_strided.rs

1use cubecl::{
2    prelude::*,
3    std::tensor::layout::{Layout, LayoutExpand},
4    {ir::DeviceProperties, prelude::barrier::Barrier},
5};
6use cubek_matmul::components::{
7    global::{
8        GlobalReaderConfig, PlaneFlowPartition,
9        memory::GlobalIterator,
10        multi_stage::LoadMaxRoundPlaneCount,
11        read::{
12            FullLoadingStrategy, LoadingJob, LoadingValidation, async_barrier::AsyncCopy,
13            async_full_strided::AsyncFullStridedLoading as MatmulStridedLoading,
14            stage::FullStageLayout,
15        },
16    },
17    stage::{StridedStageFamily, StridedStageMemory},
18};
19use cubek_matmul::definition::{MatmulElems, MatmulProblem};
20use cubek_std::{InvalidConfigError, StageIdent, tile::StridedTilingLayout};
21
22use crate::components::global::{
23    args::RuntimeArgs,
24    read::strategy::async_copy::{ASYNC_COPY_WIDTH, async_copy_from},
25};
26
27#[derive(CubeType, Clone, Copy)]
28/// Loads the content of all the stage using all planes,
29/// keeping the original layout, making each tile strided
30pub struct AsyncFullStridedLoading {}
31
32impl LoadingValidation for AsyncFullStridedLoading {
33    fn validate_with_config(
34        device_props: &DeviceProperties,
35        config: &GlobalReaderConfig,
36    ) -> Result<(), InvalidConfigError> {
37        MatmulStridedLoading::validate_with_config(device_props, config)
38    }
39
40    fn validate_with_problem(
41        problem: &MatmulProblem,
42        dtypes: &MatmulElems,
43        ident: StageIdent,
44    ) -> Result<(), InvalidConfigError> {
45        MatmulStridedLoading::validate_with_problem(problem, dtypes, ident)
46    }
47}
48
49impl LoadMaxRoundPlaneCount for AsyncFullStridedLoading {
50    fn max_round_plane_count(
51        elements_per_tile: u32,
52        tiles_per_stage: u32,
53        vector_size: VectorSize,
54        plane_dim: u32,
55        dtype: StorageType,
56    ) -> u32 {
57        MatmulStridedLoading::max_round_plane_count(
58            elements_per_tile,
59            tiles_per_stage,
60            vector_size,
61            plane_dim,
62            dtype,
63        )
64    }
65}
66
67#[cube]
68impl FullLoadingStrategy<RuntimeArgs> for AsyncFullStridedLoading {
69    type TilingLayout = StridedTilingLayout;
70    type SyncStrategy = AsyncCopy;
71    type Job<EG: Numeric, NG: Size, ES: Numeric, NS: Size> = AsyncFullStridedJob;
72    type Stage = StridedStageFamily;
73    fn new_job<EG: Numeric, NG: Size, ES: Numeric, NS: Size>(
74        runtime_args: RuntimeArgs,
75        #[comptime] config: GlobalReaderConfig,
76    ) -> Self::Job<EG, NG, ES, NS> {
77        let type_size = ES::type_size_bits().comptime();
78        let vector_size = ASYNC_COPY_WIDTH / type_size as u32;
79        let num_stage_vectors = config.smem_config.elements_per_stage() / vector_size;
80        let unit_count = config.loading_planes_count() * config.plane_dim;
81        let num_tasks_per_unit = num_stage_vectors / unit_count;
82
83        let unit_position_base = PlaneFlowPartition::new(config.plane_flow_config.partition_rule)
84            .load_index(config.input_load_flow)
85            * config.plane_dim
86            + UNIT_POS_X;
87
88        AsyncFullStridedJob {
89            unit_position_base,
90            runtime_args,
91            num_tasks_per_unit,
92            unit_count,
93            copy_vector_size: vector_size,
94        }
95    }
96}
97
98#[derive(CubeType, Clone)]
99#[expand(derive(Clone))]
100pub struct AsyncFullStridedJob {
101    unit_position_base: u32,
102    runtime_args: RuntimeArgs,
103
104    #[cube(comptime)]
105    num_tasks_per_unit: u32,
106    #[cube(comptime)]
107    unit_count: u32,
108    #[cube(comptime)]
109    copy_vector_size: u32,
110}
111
112#[cube]
113impl<EG: Numeric, NG: Size, ES: Numeric, NS: Size>
114    LoadingJob<EG, NG, ES, NS, StridedTilingLayout, AsyncCopy> for AsyncFullStridedJob
115{
116    type Stage = StridedStageFamily;
117
118    fn execute_task(
119        this: &mut Self,
120        #[comptime] task_id: u32,
121        global_iter: &GlobalIterator<Vector<EG, NG>>,
122        stage: &mut StridedStageMemory<ES, NS, StridedTilingLayout>,
123        _barrier: &Shared<Barrier>,
124        #[comptime] config: GlobalReaderConfig,
125    ) {
126        let unit_position = this.unit_position_base + task_id * this.unit_count;
127        let unit_position_abs = unit_position * this.copy_vector_size;
128
129        let layout = FullStageLayout::new(config.smem_config);
130        let view = global_iter.view();
131
132        let pos = layout.to_source_pos(unit_position_abs);
133        let stage_offset = unit_position_abs / stage.smem.vector_size() as u32;
134
135        async_copy_from(
136            view,
137            pos,
138            stage,
139            stage_offset,
140            &this.runtime_args,
141            global_iter.offset(),
142            config,
143            this.copy_vector_size,
144        );
145    }
146
147    fn task_count(this: &Self) -> comptime_type!(u32) {
148        this.num_tasks_per_unit
149    }
150}
cubek_convolution/components/global/read/strategy/async_full_strided.rs

cubek_convolution/components/global/read/strategy/
async_full_strided.rs
