baracuda_kernels/embedding/
embedding_backward.rs

1//! `embedding` BW plan — Category M.
2//!
3//! Adjoint of [`crate::embedding::EmbeddingPlan`]:
4//! `dweight[indices[n], :] += dout[n, :]` along the row axis (atomicAdd).
5//! Rows where `indices[n] == padding_idx` (or where `indices[n]` is
6//! negative / out-of-range) are skipped — no contribution.
7//!
8//! Trailblazer dtype coverage: `f32, f64`. atomicAdd is native-FP.
9//!
10//! Caller MUST zero `dweight` before launch (or pre-populate to
11//! accumulate into a running gradient).
12
13use core::ffi::c_void;
14use core::marker::PhantomData;
15
16use baracuda_cutlass::{Error, Result};
17use baracuda_driver::Stream;
18use baracuda_kernels_types::{
19    ArchSku, BackendKind, Element, ElementKind, EmbeddingKind, IndexElement, IndexElementKind,
20    KernelSku, MathPrecision, OpCategory, PlanPreference, PrecisionGuarantee, TensorMut,
21    TensorRef, Workspace,
22};
23
24use crate::indexing::gather::map_status;
25
26use super::PADDING_DISABLED;
27
28/// Descriptor for an `embedding_backward` op.
29#[derive(Copy, Clone, Debug)]
30pub struct EmbeddingBackwardDescriptor {
31    /// Vocabulary size — extent of `dweight` along axis 0.
32    pub num_embeddings: i32,
33    /// Embedding dimension — extent of `dweight` along axis 1.
34    pub embedding_dim: i32,
35    /// Number of indices — extent of `indices` and of `dout` along axis 0.
36    pub num_indices: i32,
37    /// Optional padding index. When `Some(p)`, rows where
38    /// `indices[n] == p` are skipped (no contribution).
39    pub padding_idx: Option<i32>,
40    /// Value element type.
41    pub element: ElementKind,
42}
43
44/// Args bundle for an `embedding_backward` launch.
45///
46/// Phase 11.5: `I: IndexElement` generic (`i32` or `i64`).
47pub struct EmbeddingBackwardArgs<'a, T: Element, I: IndexElement = i32> {
48    /// Upstream gradient `[N, D]`. Row-major contiguous.
49    pub dout: TensorRef<'a, T, 2>,
50    /// Index tensor `[N]` (same as FW pass).
51    pub indices: TensorRef<'a, I, 1>,
52    /// Gradient w.r.t. `weight` `[V, D]`. Caller MUST pre-zero this.
53    pub dweight: TensorMut<'a, T, 2>,
54}
55
56/// `embedding_backward` plan.
57///
58/// Adjoint of [`crate::EmbeddingPlan`]:
59/// `dweight[indices[n], :] += dout[n, :]` via atomicAdd. Rows where
60/// `indices[n] == padding_idx` (or negative / out-of-range) are skipped.
61///
62/// **When to use**: backward for [`EmbeddingPlan`](crate::EmbeddingPlan).
63///
64/// **Dtypes**: `{f32, f64}` only — atomicAdd is native-FP.
65///
66/// **Shape limits**: `dweight` is `[V, D]`, `dout` is `[N, D]`,
67/// `indices` is `[N]`, all extents non-negative.
68///
69/// **Workspace**: none. Caller MUST zero `dweight` before launch
70/// (or pre-populate to accumulate into a running gradient).
71///
72/// **Precision guarantee**: **non-deterministic** — atomicAdd
73/// ordering varies between launches.
74pub struct EmbeddingBackwardPlan<T: Element> {
75    desc: EmbeddingBackwardDescriptor,
76    sku: KernelSku,
77    _marker: PhantomData<T>,
78}
79
80impl<T: Element> EmbeddingBackwardPlan<T> {
81    /// Pick a kernel for `desc`.
82    pub fn select(
83        _stream: &Stream,
84        desc: &EmbeddingBackwardDescriptor,
85        _pref: PlanPreference,
86    ) -> Result<Self> {
87        if desc.element != T::KIND {
88            return Err(Error::Unsupported(
89                "baracuda-kernels::EmbeddingBackwardPlan: descriptor element != type parameter T",
90            ));
91        }
92        if desc.num_embeddings < 0
93            || desc.embedding_dim < 0
94            || desc.num_indices < 0
95        {
96            return Err(Error::InvalidProblem(
97                "baracuda-kernels::EmbeddingBackwardPlan: num_embeddings / embedding_dim / \
98                 num_indices must be non-negative",
99            ));
100        }
101        let supported = matches!(T::KIND, ElementKind::F32 | ElementKind::F64);
102        if !supported {
103            return Err(Error::Unsupported(
104                "baracuda-kernels::EmbeddingBackwardPlan: today only `f32`, `f64` wired \
105                 (BW uses atomicAdd)",
106            ));
107        }
108        let precision_guarantee = PrecisionGuarantee {
109            math_precision: if T::KIND == ElementKind::F64 {
110                MathPrecision::F64
111            } else {
112                MathPrecision::F32
113            },
114            accumulator: T::KIND,
115            // atomicAdd order is non-deterministic across launches.
116            bit_stable_on_same_hardware: false,
117            deterministic: false,
118        };
119        let sku = KernelSku {
120            category: OpCategory::Embedding,
121            op: EmbeddingKind::EmbeddingBackward as u16,
122            element: T::KIND,
123            aux_element: Some(ElementKind::I32),
124            layout: None,
125            epilogue: None,
126            arch: ArchSku::Sm80,
127            backend: BackendKind::Bespoke,
128            precision_guarantee,
129        };
130        Ok(Self {
131            desc: *desc,
132            sku,
133            _marker: PhantomData,
134        })
135    }
136
137    /// Validate args.
138    pub fn can_implement<I: IndexElement>(&self, args: &EmbeddingBackwardArgs<'_, T, I>) -> Result<()> {
139        if args.dout.shape[0] != self.desc.num_indices
140            || args.dout.shape[1] != self.desc.embedding_dim
141        {
142            return Err(Error::InvalidProblem(
143                "baracuda-kernels::EmbeddingBackwardPlan: dout shape must be \
144                 [num_indices, embedding_dim]",
145            ));
146        }
147        if args.indices.shape[0] != self.desc.num_indices {
148            return Err(Error::InvalidProblem(
149                "baracuda-kernels::EmbeddingBackwardPlan: indices.shape[0] mismatch with descriptor",
150            ));
151        }
152        if args.dweight.shape[0] != self.desc.num_embeddings
153            || args.dweight.shape[1] != self.desc.embedding_dim
154        {
155            return Err(Error::InvalidProblem(
156                "baracuda-kernels::EmbeddingBackwardPlan: dweight shape must be \
157                 [num_embeddings, embedding_dim]",
158            ));
159        }
160        let dout_len = args.dout.data.len() as i64;
161        let idx_len = args.indices.data.len() as i64;
162        let dw_len = args.dweight.data.len() as i64;
163        let dout_numel = args.dout.numel();
164        let idx_numel = args.indices.numel();
165        let dw_numel = args.dweight.numel();
166        if dout_len < dout_numel {
167            return Err(Error::BufferTooSmall {
168                needed: dout_numel as usize,
169                got: dout_len as usize,
170            });
171        }
172        if idx_len < idx_numel {
173            return Err(Error::BufferTooSmall {
174                needed: idx_numel as usize,
175                got: idx_len as usize,
176            });
177        }
178        if dw_len < dw_numel {
179            return Err(Error::BufferTooSmall {
180                needed: dw_numel as usize,
181                got: dw_len as usize,
182            });
183        }
184        Ok(())
185    }
186
187    /// Workspace size in bytes (zero).
188    #[inline]
189    pub fn workspace_size(&self) -> usize {
190        0
191    }
192
193    /// Identity of the kernel this plan picked.
194    #[inline]
195    pub fn sku(&self) -> KernelSku {
196        self.sku
197    }
198
199    /// Numerical guarantees for this plan's kernel.
200    #[inline]
201    pub fn precision_guarantee(&self) -> PrecisionGuarantee {
202        self.sku.precision_guarantee
203    }
204
205    /// Launch.
206    ///
207    /// Phase 11.5: generic over `I: IndexElement`.
208    pub fn run<I: IndexElement>(
209        &self,
210        stream: &Stream,
211        _workspace: Workspace<'_>,
212        args: EmbeddingBackwardArgs<'_, T, I>,
213    ) -> Result<()> {
214        self.can_implement(&args)?;
215        let num_indices = self.desc.num_indices as i64;
216        if num_indices == 0 || self.desc.embedding_dim == 0 {
217            return Ok(());
218        }
219        let dout_ptr = args.dout.data.as_raw().0 as *const c_void;
220        let idx_ptr = args.indices.data.as_raw().0 as *const c_void;
221        let dweight_ptr = args.dweight.data.as_raw().0 as *mut c_void;
222        let stream_ptr = stream.as_raw() as *mut c_void;
223        // Phase 11.5: padding_idx widens to i64 across FFI.
224        let padding_idx: i64 = self.desc.padding_idx.unwrap_or(PADDING_DISABLED) as i64;
225
226        let status = match (T::KIND, I::KIND) {
227            (ElementKind::F32, IndexElementKind::I32) => unsafe {
228                baracuda_kernels_sys::baracuda_kernels_embedding_backward_f32_run(
229                    num_indices, self.desc.num_embeddings, self.desc.embedding_dim,
230                    padding_idx, dout_ptr, idx_ptr, dweight_ptr,
231                    core::ptr::null_mut(), 0, stream_ptr,
232                )
233            },
234            (ElementKind::F64, IndexElementKind::I32) => unsafe {
235                baracuda_kernels_sys::baracuda_kernels_embedding_backward_f64_run(
236                    num_indices, self.desc.num_embeddings, self.desc.embedding_dim,
237                    padding_idx, dout_ptr, idx_ptr, dweight_ptr,
238                    core::ptr::null_mut(), 0, stream_ptr,
239                )
240            },
241            (ElementKind::F32, IndexElementKind::I64) => unsafe {
242                baracuda_kernels_sys::baracuda_kernels_embedding_backward_i64idx_f32_run(
243                    num_indices, self.desc.num_embeddings, self.desc.embedding_dim,
244                    padding_idx, dout_ptr, idx_ptr, dweight_ptr,
245                    core::ptr::null_mut(), 0, stream_ptr,
246                )
247            },
248            (ElementKind::F64, IndexElementKind::I64) => unsafe {
249                baracuda_kernels_sys::baracuda_kernels_embedding_backward_i64idx_f64_run(
250                    num_indices, self.desc.num_embeddings, self.desc.embedding_dim,
251                    padding_idx, dout_ptr, idx_ptr, dweight_ptr,
252                    core::ptr::null_mut(), 0, stream_ptr,
253                )
254            },
255            _ => {
256                return Err(Error::Unsupported(
257                    "baracuda-kernels::EmbeddingBackwardPlan::run reached an unimplemented dtype \
258                     — select() should have caught this",
259                ));
260            }
261        };
262        map_status(status)
263    }
264}
baracuda_kernels/embedding/embedding_backward.rs

baracuda_kernels/embedding/
embedding_backward.rs
