cubecl_convolution/kernels/layered/algorithm/
simple_tma.rs

1use cubecl_core::server::LaunchError;
2use cubecl_core::{
3    CubeCount, Runtime, client::ComputeClient, ir::StorageType, prelude::TensorHandleRef,
4};
5use cubecl_matmul::components::{
6    InvalidConfigError, global::args::TensorMapArgs, stage::PlaneMatmulFamily,
7    tile::TileMatmulFamily,
8};
9use cubecl_matmul::components::{
10    MatmulElems, MatmulSelection, MatmulSetupError, stage::StridedStageFamily, tile::io::Strided,
11};
12use cubecl_matmul::components::{MatmulLineSizes, stage::NumStages};
13use cubecl_std::{
14    CubeOption,
15    tensor::{TensorHandle, into_contiguous_pitched},
16};
17use std::marker::PhantomData;
18
19use crate::components::{
20    ConvolutionProblem, Dimensionality, convolution_matmul_selection,
21    global::single_stage::tma::SimpleTmaConvolutionFamily,
22};
23
24use super::Algorithm;
25
26pub const TMA_STRIDE_ALIGN: usize = 16;
27
28/// Cmma convolution
29pub struct SimpleTmaConvAlgorithm<TMM: TileMatmulFamily> {
30    _tmm: PhantomData<TMM>,
31}
32
33impl<
34    TMM: TileMatmulFamily<
35            LhsTile = Strided,
36            RhsTile = Strided,
37            AccTile = CubeOption<Strided>,
38            OutTile = Strided,
39        >,
40> Algorithm for SimpleTmaConvAlgorithm<TMM>
41{
42    type TileMatmul = TMM;
43    type StageMatmul = PlaneMatmulFamily<
44        Self::TileMatmul,
45        StridedStageFamily,
46        StridedStageFamily,
47        Option<StridedStageFamily>,
48    >;
49    type GlobalConvolution = SimpleTmaConvolutionFamily<Self::StageMatmul>;
50
51    type Args = TensorMapArgs;
52
53    fn cube_count(selection: &MatmulSelection, problem: &ConvolutionProblem) -> CubeCount {
54        let m_stage = selection.tiling_scheme.elements_per_stage_along_m();
55        let n_stage = selection.tiling_scheme.elements_per_stage_along_n();
56        let cubes_needed_m = (problem.m as u32).div_ceil(m_stage);
57        let cubes_needed_n = (problem.n as u32).div_ceil(n_stage);
58
59        CubeCount::Static(cubes_needed_m, cubes_needed_n, 1)
60    }
61
62    fn into_tensor_handle<R: Runtime>(
63        client: &ComputeClient<R>,
64        handle: &TensorHandleRef<'_, R>,
65        dtype: StorageType,
66    ) -> Result<TensorHandle<R>, LaunchError> {
67        into_tensor_handle_tma(client, handle, dtype)
68    }
69
70    // TODO this is not the same as tma stages, it's stages in the sense of double buffering in matmul
71    fn num_stages() -> NumStages {
72        (1, 1).into()
73    }
74
75    fn selection<R: Runtime>(
76        client: &ComputeClient<R>,
77        problem: &ConvolutionProblem,
78        plane_dim: u32,
79        line_sizes: &MatmulLineSizes,
80        dtypes: &mut MatmulElems,
81    ) -> Result<MatmulSelection, MatmulSetupError> {
82        Ok(convolution_matmul_selection::<TMM, R>(
83            client, problem, plane_dim, false, line_sizes, dtypes,
84        )?)
85    }
86}
87
88pub(crate) fn into_tensor_handle_tma<R: Runtime>(
89    client: &ComputeClient<R>,
90    handle: &TensorHandleRef<'_, R>,
91    dtype: StorageType,
92) -> Result<TensorHandle<R>, LaunchError> {
93    let handle = if has_valid_layout(handle) {
94        TensorHandle::from_ref(handle, dtype)
95    } else {
96        into_contiguous_pitched(client, handle, dtype)?
97    };
98    Ok(handle)
99}
100
101pub(crate) fn has_valid_layout<R: Runtime>(handle: &TensorHandleRef<'_, R>) -> bool {
102    let stride_align = TMA_STRIDE_ALIGN / handle.elem_size;
103    let rank = handle.shape.len();
104    let dim_c = rank - 1;
105
106    let aligned = handle.strides[..dim_c]
107        .iter()
108        .all(|stride| stride % stride_align == 0);
109
110    let valid_layout = handle.strides[dim_c] == 1;
111
112    valid_layout && aligned
113}
114
115pub(crate) fn check_problem_tma(problem: &ConvolutionProblem) -> Result<(), InvalidConfigError> {
116    fn check_range(
117        value: isize,
118        name: impl FnOnce() -> String,
119        min: isize,
120        max: isize,
121    ) -> Result<(), InvalidConfigError> {
122        if value < min || value > max {
123            let name = name();
124            Err(Box::new(format!(
125                "value {name} outside of valid range ({min}, {max})"
126            )))
127        } else {
128            Ok(())
129        }
130    }
131
132    let (corner_min, corner_max) = match problem.dimensionality {
133        Dimensionality::Dim1 => (-(2isize.pow(15)), 2isize.pow(15) - 1),
134        Dimensionality::Dim2 => (-(2isize.pow(7)), 2isize.pow(7) - 1),
135        Dimensionality::Dim3 => (-(2isize.pow(4)), 2isize.pow(4) - 1),
136    };
137
138    let corner = calculate_upper_corner(&problem.padding, &problem.kernel_size, &problem.dilation);
139    for (i, offs) in corner.iter().enumerate() {
140        check_range(
141            *offs as isize,
142            || format!("corner[{i}]"),
143            corner_min,
144            corner_max,
145        )?;
146    }
147
148    let offset_max = match problem.dimensionality {
149        Dimensionality::Dim1 => 2isize.pow(16) - 1,
150        Dimensionality::Dim2 => 2isize.pow(8) - 1,
151        Dimensionality::Dim3 => 2isize.pow(5) - 1,
152    };
153
154    for i in 0..problem.kernel_size.len() {
155        let offset = (problem.kernel_size[i] - 1) * problem.dilation[i];
156        check_range(
157            offset as isize,
158            || format!("kernel size {i}"),
159            0,
160            offset_max,
161        )?;
162        check_range(problem.stride[i] as isize, || format!("stride[{i}]"), 1, 8)?;
163    }
164
165    Ok(())
166}
167
168pub fn calculate_lower_corner(padding: &[i32]) -> Vec<i32> {
169    padding.iter().map(|padding| -*padding).collect()
170}
171
172pub fn calculate_upper_corner(padding: &[i32], kernel_size: &[u32], dilation: &[u32]) -> Vec<i32> {
173    padding
174        .iter()
175        .zip(kernel_size)
176        .zip(dilation)
177        .map(|((padding, kernel_size), dilation)| {
178            *padding - (*kernel_size - 1) as i32 * *dilation as i32
179        })
180        .collect()
181}
cubecl_convolution/kernels/layered/algorithm/simple_tma.rs

cubecl_convolution/kernels/layered/algorithm/
simple_tma.rs
