baracuda_kernels/softmax/
axis_backward.rs

1//! Softmax backward plan.
2//!
3//! **Formulas**:
4//! - `Softmax` BW:    `dx[k] = y[k] · (dy[k] - Σ_j y[j] · dy[j])`
5//! - `LogSoftmax` BW: `dx[k] = dy[k] - exp(y[k]) · Σ_j dy[j]`
6//!
7//! Both reference the **saved forward output** `y` — the BW formula has
8//! no dependence on the forward input `x` once `y` is known.
9//!
10//! **When to use**: autograd backward for [`SoftmaxPlan`](super::SoftmaxPlan).
11//! Caller saves `y` from the FW pass and feeds it as `args.y`.
12//!
13//! **Dtypes / shape**: `{Softmax, LogSoftmax} × {f32, f16, bf16, f64}`,
14//! tensor rank `1..=8`. f16 / bf16 reduce in f32 (FP detour) then cast
15//! back.
16//!
17//! **Workspace**: none.
18//!
19//! **Precision**: deterministic, bit-stable on the same hardware
20//! (two-pass per-row scan; no atomic-add).
21
22use core::ffi::c_void;
23use core::marker::PhantomData;
24
25use baracuda_cutlass::{Error, Result};
26use baracuda_driver::Stream;
27use baracuda_kernels_types::{
28    ArchSku, BackendKind, Element, ElementKind, KernelSku, MathPrecision, OpCategory,
29    PlanPreference, PrecisionGuarantee, SoftmaxKind, TensorMut, TensorRef, Workspace,
30};
31
32/// Descriptor for a softmax-family BW op.
33#[derive(Copy, Clone, Debug)]
34pub struct SoftmaxBackwardDescriptor<const N: usize> {
35    /// Which softmax variant this is the BW of.
36    pub kind: SoftmaxKind,
37    /// Tensor shape (shared by dy / y / dx).
38    pub input_shape: [i32; N],
39    /// Forward softmax axis.
40    pub softmax_axis: u8,
41    /// Element type.
42    pub element: ElementKind,
43}
44
45/// Args bundle for a softmax BW launch.
46///
47/// `y` is the SAVED forward output. Required by all softmax-family BW
48/// kernels (the gradient formula references it).
49pub struct SoftmaxBackwardArgs<'a, T: Element, const N: usize> {
50    /// Upstream gradient.
51    pub dy: TensorRef<'a, T, N>,
52    /// Saved forward output.
53    pub y: TensorRef<'a, T, N>,
54    /// Gradient w.r.t. the forward input.
55    pub dx: TensorMut<'a, T, N>,
56}
57
58/// Softmax backward plan — see the module-level docs for formulas,
59/// dtypes, workspace, and precision guarantees.
60///
61/// `T: Element` is the element type (`f32` / `f64` / `f16` / `bf16`).
62/// `const N: usize` is the tensor rank (1..=8).
63pub struct SoftmaxBackwardPlan<T: Element, const N: usize> {
64    desc: SoftmaxBackwardDescriptor<N>,
65    sku: KernelSku,
66    _marker: PhantomData<T>,
67}
68
69impl<T: Element, const N: usize> SoftmaxBackwardPlan<T, N> {
70    /// Pick a kernel for `desc`. Validates `softmax_axis < N`, the dtype
71    /// is in the wired FP family, and tensor rank ≤ 8. Returns
72    /// [`Error::Unsupported`] for cells outside the matrix and
73    /// [`Error::InvalidProblem`] for malformed shapes / axes.
74    pub fn select(
75        _stream: &Stream,
76        desc: &SoftmaxBackwardDescriptor<N>,
77        _pref: PlanPreference,
78    ) -> Result<Self> {
79        if desc.element != T::KIND {
80            return Err(Error::Unsupported(
81                "baracuda-kernels::SoftmaxBackwardPlan: descriptor element != T",
82            ));
83        }
84        if (desc.softmax_axis as usize) >= N {
85            return Err(Error::InvalidProblem(
86                "baracuda-kernels::SoftmaxBackwardPlan: softmax_axis out of range for rank N",
87            ));
88        }
89        for &d in desc.input_shape.iter() {
90            if d < 0 {
91                return Err(Error::InvalidProblem(
92                    "baracuda-kernels::SoftmaxBackwardPlan: shape dims must be non-negative",
93                ));
94            }
95        }
96        if N > 8 {
97            return Err(Error::Unsupported(
98                "baracuda-kernels::SoftmaxBackwardPlan: tensor rank > 8 not supported",
99            ));
100        }
101        let dtype_in_fp_family = matches!(
102            T::KIND,
103            ElementKind::F32 | ElementKind::F16 | ElementKind::Bf16 | ElementKind::F64
104        );
105        let kind_supported = matches!(desc.kind, SoftmaxKind::Softmax | SoftmaxKind::LogSoftmax);
106        if !kind_supported || !dtype_in_fp_family {
107            return Err(Error::Unsupported(
108                "baracuda-kernels::SoftmaxBackwardPlan: wired today: \
109                 `{Softmax, LogSoftmax} × {f32, f16, bf16, f64}`",
110            ));
111        }
112
113        let precision_guarantee = PrecisionGuarantee {
114            math_precision: MathPrecision::F32,
115            accumulator: ElementKind::F32,
116            bit_stable_on_same_hardware: true,
117            deterministic: true,
118        };
119        let sku = KernelSku {
120            category: OpCategory::Softmax,
121            op: desc.kind as u16,
122            element: T::KIND,
123            aux_element: None,
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(&self, args: &SoftmaxBackwardArgs<'_, T, N>) -> Result<()> {
139        if args.dy.shape != self.desc.input_shape {
140            return Err(Error::InvalidProblem(
141                "baracuda-kernels::SoftmaxBackwardPlan: dy shape mismatch",
142            ));
143        }
144        if args.y.shape != self.desc.input_shape {
145            return Err(Error::InvalidProblem(
146                "baracuda-kernels::SoftmaxBackwardPlan: y shape mismatch",
147            ));
148        }
149        if args.dx.shape != self.desc.input_shape {
150            return Err(Error::InvalidProblem(
151                "baracuda-kernels::SoftmaxBackwardPlan: dx shape mismatch",
152            ));
153        }
154        let numel = args.dx.numel();
155        let dy_len = args.dy.data.len() as i64;
156        let y_len = args.y.data.len() as i64;
157        let dx_len = args.dx.data.len() as i64;
158        if dy_len < numel || y_len < numel || dx_len < numel {
159            return Err(Error::BufferTooSmall {
160                needed: numel as usize,
161                got: dy_len.min(y_len).min(dx_len) as usize,
162            });
163        }
164        Ok(())
165    }
166
167    /// Workspace size in bytes. Always zero.
168    #[inline]
169    pub fn workspace_size(&self) -> usize {
170        0
171    }
172    /// Identity of the kernel this plan picked.
173    #[inline]
174    pub fn sku(&self) -> KernelSku {
175        self.sku
176    }
177    /// Numerical guarantees for this plan's kernel — deterministic,
178    /// bit-stable on the same hardware, f32 accumulator for f16 / bf16
179    /// inputs (FP detour).
180    #[inline]
181    pub fn precision_guarantee(&self) -> PrecisionGuarantee {
182        self.sku.precision_guarantee
183    }
184
185    /// Launch the kernel against `args`. Calls `can_implement` first;
186    /// returns `Ok(())` for empty tensors.
187    pub fn run(
188        &self,
189        stream: &Stream,
190        _workspace: Workspace<'_>,
191        args: SoftmaxBackwardArgs<'_, T, N>,
192    ) -> Result<()> {
193        self.can_implement(&args)?;
194        let numel = args.dx.numel();
195        if numel == 0 {
196            return Ok(());
197        }
198        let dy_ptr = args.dy.data.as_raw().0 as *const c_void;
199        let y_ptr = args.y.data.as_raw().0 as *const c_void;
200        let dx_ptr = args.dx.data.as_raw().0 as *mut c_void;
201        let stream_ptr = stream.as_raw() as *mut c_void;
202
203        let axis = self.desc.softmax_axis as usize;
204        let shape = self.desc.input_shape;
205        let stride_dy = args.dy.stride;
206        let stride_y = args.y.stride;
207        let stride_dx = args.dx.stride;
208        let rank = N as i32;
209        let extent = shape[axis];
210        let stride_dy_axis = stride_dy[axis];
211        let stride_y_axis = stride_y[axis];
212
213        macro_rules! dispatch {
214            ($sym:ident) => {
215                unsafe {
216                    baracuda_kernels_sys::$sym(
217                        numel,
218                        rank,
219                        shape.as_ptr(),
220                        stride_dy.as_ptr(),
221                        stride_y.as_ptr(),
222                        stride_dx.as_ptr(),
223                        axis as i32,
224                        extent,
225                        stride_dy_axis,
226                        stride_y_axis,
227                        dy_ptr,
228                        y_ptr,
229                        dx_ptr,
230                        core::ptr::null_mut(),
231                        0,
232                        stream_ptr,
233                    )
234                }
235            };
236        }
237
238        let status = match (self.desc.kind, T::KIND) {
239            (SoftmaxKind::Softmax, ElementKind::F32) => {
240                dispatch!(baracuda_kernels_softmax_backward_f32_run)
241            }
242            (SoftmaxKind::Softmax, ElementKind::F16) => {
243                dispatch!(baracuda_kernels_softmax_backward_f16_run)
244            }
245            (SoftmaxKind::Softmax, ElementKind::Bf16) => {
246                dispatch!(baracuda_kernels_softmax_backward_bf16_run)
247            }
248            (SoftmaxKind::Softmax, ElementKind::F64) => {
249                dispatch!(baracuda_kernels_softmax_backward_f64_run)
250            }
251            (SoftmaxKind::LogSoftmax, ElementKind::F32) => {
252                dispatch!(baracuda_kernels_log_softmax_backward_f32_run)
253            }
254            (SoftmaxKind::LogSoftmax, ElementKind::F16) => {
255                dispatch!(baracuda_kernels_log_softmax_backward_f16_run)
256            }
257            (SoftmaxKind::LogSoftmax, ElementKind::Bf16) => {
258                dispatch!(baracuda_kernels_log_softmax_backward_bf16_run)
259            }
260            (SoftmaxKind::LogSoftmax, ElementKind::F64) => {
261                dispatch!(baracuda_kernels_log_softmax_backward_f64_run)
262            }
263            _ => {
264                return Err(Error::Unsupported(
265                    "baracuda-kernels::SoftmaxBackwardPlan::run reached an unimplemented \
266                     (kind, dtype) pair — select() should have caught this",
267                ));
268            }
269        };
270        map_status(status)
271    }
272}
273
274fn map_status(code: i32) -> Result<()> {
275    match code {
276        0 => Ok(()),
277        1 => Err(Error::MisalignedOperand),
278        2 => Err(Error::InvalidProblem(
279            "baracuda-kernels-sys reported invalid problem",
280        )),
281        3 => Err(Error::Unsupported(
282            "baracuda-kernels-sys reported unsupported configuration",
283        )),
284        4 => Err(Error::WorkspaceTooSmall { needed: 0, got: 0 }),
285        n => Err(Error::CutlassInternal(n)),
286    }
287}
baracuda_kernels/softmax/axis_backward.rs

baracuda_kernels/softmax/
axis_backward.rs
