baracuda_kernels/image/
grid_sample.rs

1//! `grid_sample` FW plan — Category T (2-D).
2//!
3//! Sample an NCHW input at arbitrary normalized (x, y) coordinates
4//! supplied via `grid: [N, OH, OW, 2]`. Trailblazer config matches
5//! PyTorch defaults: `mode='bilinear'`, `padding_mode='zeros'`,
6//! `align_corners=false`.
7//!
8//! Trailblazer dtype coverage: `f32, f64`.
9
10use core::ffi::c_void;
11use core::marker::PhantomData;
12
13use baracuda_cutlass::{Error, Result};
14use baracuda_driver::Stream;
15use baracuda_kernels_types::{
16    ArchSku, BackendKind, Element, ElementKind, ImageKind, KernelSku, MathPrecision, OpCategory,
17    PlanPreference, PrecisionGuarantee, TensorMut, TensorRef, Workspace,
18};
19
20use super::map_status;
21
22/// Descriptor for a `grid_sample` op.
23#[derive(Copy, Clone, Debug)]
24pub struct GridSampleDescriptor {
25    /// Batch.
26    pub n: i32,
27    /// Channels.
28    pub c: i32,
29    /// Input height.
30    pub ih: i32,
31    /// Input width.
32    pub iw: i32,
33    /// Output height.
34    pub oh: i32,
35    /// Output width.
36    pub ow: i32,
37    /// Value element type.
38    pub element: ElementKind,
39}
40
41/// Args bundle for a `grid_sample` launch.
42pub struct GridSampleArgs<'a, T: Element> {
43    /// Input `[N, C, IH, IW]`. NCHW row-major contiguous.
44    pub input: TensorRef<'a, T, 4>,
45    /// Sampling grid `[N, OH, OW, 2]` — (x, y) normalized in [-1, 1].
46    pub grid: TensorRef<'a, T, 4>,
47    /// Output `[N, C, OH, OW]`.
48    pub output: TensorMut<'a, T, 4>,
49}
50
51/// `grid_sample` plan.
52///
53/// Sample an NCHW input at arbitrary normalized `(x, y)` coordinates
54/// supplied via `grid: [N, OH, OW, 2]` (PyTorch `F.grid_sample`).
55/// Trailblazer config matches PyTorch defaults: `mode='bilinear'`,
56/// `padding_mode='zeros'`, `align_corners=false`.
57///
58/// **When to use**: forward 2-D grid sample. Pair with
59/// [`GridSampleBackwardPlan`](crate::GridSampleBackwardPlan).
60///
61/// **Dtypes**: `{f32, f64}`.
62///
63/// **Shape limits**: rank-4 NCHW input; rank-4 grid `[N, OH, OW, 2]`;
64/// output `[N, C, OH, OW]`.
65///
66/// **Workspace**: none.
67///
68/// **Precision guarantee**: deterministic, bit-stable.
69pub struct GridSamplePlan<T: Element> {
70    desc: GridSampleDescriptor,
71    sku: KernelSku,
72    _marker: PhantomData<T>,
73}
74
75impl<T: Element> GridSamplePlan<T> {
76    /// Pick a kernel.
77    pub fn select(
78        _stream: &Stream,
79        desc: &GridSampleDescriptor,
80        _pref: PlanPreference,
81    ) -> Result<Self> {
82        if desc.element != T::KIND {
83            return Err(Error::Unsupported(
84                "baracuda-kernels::GridSamplePlan: descriptor element != T",
85            ));
86        }
87        if desc.n < 0 || desc.c < 0 || desc.ih < 0 || desc.iw < 0 || desc.oh < 0 || desc.ow < 0 {
88            return Err(Error::InvalidProblem(
89                "baracuda-kernels::GridSamplePlan: all extents must be non-negative",
90            ));
91        }
92        if !matches!(T::KIND, ElementKind::F32 | ElementKind::F64) {
93            return Err(Error::Unsupported(
94                "baracuda-kernels::GridSamplePlan: only `f32`, `f64` wired",
95            ));
96        }
97        let precision_guarantee = PrecisionGuarantee {
98            math_precision: if T::KIND == ElementKind::F64 {
99                MathPrecision::F64
100            } else {
101                MathPrecision::F32
102            },
103            accumulator: T::KIND,
104            bit_stable_on_same_hardware: true,
105            deterministic: true,
106        };
107        let sku = KernelSku {
108            category: OpCategory::Image,
109            op: ImageKind::GridSample2d as u16,
110            element: T::KIND,
111            aux_element: None,
112            layout: None,
113            epilogue: None,
114            arch: ArchSku::Sm80,
115            backend: BackendKind::Bespoke,
116            precision_guarantee,
117        };
118        Ok(Self {
119            desc: *desc,
120            sku,
121            _marker: PhantomData,
122        })
123    }
124
125    /// Validate args.
126    pub fn can_implement(&self, args: &GridSampleArgs<'_, T>) -> Result<()> {
127        if args.input.shape != [self.desc.n, self.desc.c, self.desc.ih, self.desc.iw] {
128            return Err(Error::InvalidProblem(
129                "baracuda-kernels::GridSamplePlan: input shape mismatch",
130            ));
131        }
132        if args.grid.shape != [self.desc.n, self.desc.oh, self.desc.ow, 2] {
133            return Err(Error::InvalidProblem(
134                "baracuda-kernels::GridSamplePlan: grid shape must be [N, OH, OW, 2]",
135            ));
136        }
137        if args.output.shape != [self.desc.n, self.desc.c, self.desc.oh, self.desc.ow] {
138            return Err(Error::InvalidProblem(
139                "baracuda-kernels::GridSamplePlan: output shape mismatch",
140            ));
141        }
142        Ok(())
143    }
144
145    /// Workspace (zero).
146    #[inline]
147    pub fn workspace_size(&self) -> usize {
148        0
149    }
150
151    /// Identity.
152    #[inline]
153    pub fn sku(&self) -> KernelSku {
154        self.sku
155    }
156
157    /// Numerical guarantees.
158    #[inline]
159    pub fn precision_guarantee(&self) -> PrecisionGuarantee {
160        self.sku.precision_guarantee
161    }
162
163    /// Launch.
164    pub fn run(
165        &self,
166        stream: &Stream,
167        _workspace: Workspace<'_>,
168        args: GridSampleArgs<'_, T>,
169    ) -> Result<()> {
170        self.can_implement(&args)?;
171        if args.output.numel() == 0 {
172            return Ok(());
173        }
174        let input_ptr = args.input.data.as_raw().0 as *const c_void;
175        let grid_ptr = args.grid.data.as_raw().0 as *const c_void;
176        let out_ptr = args.output.data.as_raw().0 as *mut c_void;
177        let stream_ptr = stream.as_raw() as *mut c_void;
178        let status = match T::KIND {
179            ElementKind::F32 => unsafe {
180                baracuda_kernels_sys::baracuda_kernels_grid_sample_2d_f32_run(
181                    self.desc.n, self.desc.c, self.desc.ih, self.desc.iw,
182                    self.desc.oh, self.desc.ow,
183                    input_ptr, grid_ptr, out_ptr,
184                    core::ptr::null_mut(), 0, stream_ptr,
185                )
186            },
187            ElementKind::F64 => unsafe {
188                baracuda_kernels_sys::baracuda_kernels_grid_sample_2d_f64_run(
189                    self.desc.n, self.desc.c, self.desc.ih, self.desc.iw,
190                    self.desc.oh, self.desc.ow,
191                    input_ptr, grid_ptr, out_ptr,
192                    core::ptr::null_mut(), 0, stream_ptr,
193                )
194            },
195            _ => {
196                return Err(Error::Unsupported(
197                    "baracuda-kernels::GridSamplePlan::run reached unimplemented dtype",
198                ));
199            }
200        };
201        map_status(status)
202    }
203}
baracuda_kernels/image/grid_sample.rs

baracuda_kernels/image/
grid_sample.rs
