baracuda_kernels/reduce/
bool_axis.rs

1//! Single-axis Any / All reductions. Heterogeneous output dtype: input
2//! is `T: Element`, output is always [`Bool`] (PyTorch convention —
3//! `u8` storage, 0 = false, 1 = true).
4//!
5//! Sibling of [`crate::ReducePlan`] (same-dtype-in-out) and
6//! [`crate::ArgReducePlan`] (i64-out). Lives in its own plan shape
7//! because of the heterogeneous output. PyTorch parity:
8//! `torch.any(x, dim=k)` / `torch.all(x, dim=k)`.
9//!
10//! Wired matrix: `{Any, All} × {f32, f16, bf16, f64, i32, i64, Bool}`
11//! — 14 SKUs. NaN is truthy (`NaN != 0` is true — matches PyTorch /
12//! IEEE 754). Non-differentiable; no backward plan ships.
13
14use core::ffi::c_void;
15use core::marker::PhantomData;
16
17use baracuda_cutlass::{Error, Result};
18use baracuda_driver::Stream;
19use baracuda_kernels_types::{
20    ArchSku, BackendKind, Bool, Element, ElementKind, KernelSku, MathPrecision, OpCategory,
21    PlanPreference, PrecisionGuarantee, ReduceKind, TensorMut, TensorRef, Workspace,
22};
23
24/// Descriptor for an any / all axis reduction.
25#[derive(Copy, Clone, Debug)]
26pub struct BoolReduceDescriptor<const N: usize> {
27    /// Which reduction (must be [`ReduceKind::Any`] or [`ReduceKind::All`]).
28    pub kind: ReduceKind,
29    /// Input tensor shape.
30    pub input_shape: [i32; N],
31    /// Axis to reduce along.
32    pub reduce_axis: u8,
33    /// Input element type.
34    pub element: ElementKind,
35}
36
37impl<const N: usize> BoolReduceDescriptor<N> {
38    /// Output shape: input shape with reduce axis = 1 (keepdim).
39    pub fn output_shape(&self) -> [i32; N] {
40        let mut out = self.input_shape;
41        out[self.reduce_axis as usize] = 1;
42        out
43    }
44}
45
46/// Args bundle. Output is always [`Bool`] regardless of input dtype.
47pub struct BoolReduceArgs<'a, T: Element, const N: usize> {
48    /// Input tensor.
49    pub x: TensorRef<'a, T, N>,
50    /// Output tensor — always `Bool` (u8 storage, 0/1).
51    pub y: TensorMut<'a, Bool, N>,
52}
53
54/// Plan for an any / all axis reduction.
55pub struct BoolReducePlan<T: Element, const N: usize> {
56    desc: BoolReduceDescriptor<N>,
57    sku: KernelSku,
58    _marker: PhantomData<T>,
59}
60
61impl<T: Element, const N: usize> BoolReducePlan<T, N> {
62    /// Pick a kernel for `desc`.
63    pub fn select(
64        _stream: &Stream,
65        desc: &BoolReduceDescriptor<N>,
66        _pref: PlanPreference,
67    ) -> Result<Self> {
68        if desc.element != T::KIND {
69            return Err(Error::Unsupported(
70                "baracuda-kernels::BoolReducePlan: descriptor element != type parameter T",
71            ));
72        }
73        if (desc.reduce_axis as usize) >= N {
74            return Err(Error::InvalidProblem(
75                "baracuda-kernels::BoolReducePlan: reduce_axis must be < rank",
76            ));
77        }
78        for &d in desc.input_shape.iter() {
79            if d < 0 {
80                return Err(Error::InvalidProblem(
81                    "baracuda-kernels::BoolReducePlan: input_shape dims must be non-negative",
82                ));
83            }
84        }
85        let kind_in_scope = matches!(desc.kind, ReduceKind::Any | ReduceKind::All);
86        if !kind_in_scope {
87            return Err(Error::Unsupported(
88                "baracuda-kernels::BoolReducePlan: kind must be Any or All",
89            ));
90        }
91        let dtype_in_scope = matches!(
92            T::KIND,
93            ElementKind::F32
94                | ElementKind::F16
95                | ElementKind::Bf16
96                | ElementKind::F64
97                | ElementKind::I32
98                | ElementKind::I64
99                | ElementKind::Bool
100        );
101        if !dtype_in_scope {
102            return Err(Error::Unsupported(
103                "baracuda-kernels::BoolReducePlan: supported input dtypes are \
104                 {f32, f16, bf16, f64, i32, i64, Bool}",
105            ));
106        }
107        // Any / All do an integer-style OR / AND on a bool accumulator;
108        // no FP math, so the kernel is bit-stable on the same hardware.
109        let precision_guarantee = PrecisionGuarantee {
110            math_precision: MathPrecision::F32,
111            accumulator: ElementKind::Bool,
112            bit_stable_on_same_hardware: true,
113            deterministic: true,
114        };
115        let sku = KernelSku {
116            category: OpCategory::Reduction,
117            op: desc.kind as u16,
118            element: T::KIND,
119            // Output dtype is Bool — store it in the aux_element slot
120            // for telemetry; the op-discriminant + plan-shape tag is
121            // the authoritative "this is the Bool-out plan family".
122            aux_element: Some(ElementKind::Bool),
123            layout: None,
124            epilogue: None,
125            arch: ArchSku::Sm80,
126            backend: BackendKind::Bespoke,
127            precision_guarantee,
128        };
129        Ok(Self {
130            desc: *desc,
131            sku,
132            _marker: PhantomData,
133        })
134    }
135
136    /// Validate args.
137    pub fn can_implement(&self, args: &BoolReduceArgs<'_, T, N>) -> Result<()> {
138        if args.x.shape != self.desc.input_shape {
139            return Err(Error::InvalidProblem(
140                "baracuda-kernels::BoolReducePlan: X shape mismatch with descriptor",
141            ));
142        }
143        let expected_out = self.desc.output_shape();
144        if args.y.shape != expected_out {
145            return Err(Error::InvalidProblem(
146                "baracuda-kernels::BoolReducePlan: Y shape mismatch with derived output \
147                 shape (input shape with reduce_axis collapsed to 1)",
148            ));
149        }
150        if N > 8 {
151            return Err(Error::Unsupported(
152                "baracuda-kernels::BoolReducePlan: tensor rank > 8 not supported",
153            ));
154        }
155        let y_numel = args.y.numel();
156        let x_numel = args.x.numel();
157        let x_len = args.x.data.len() as i64;
158        let y_len = args.y.data.len() as i64;
159        if y_len < y_numel {
160            return Err(Error::BufferTooSmall {
161                needed: y_numel as usize,
162                got: y_len as usize,
163            });
164        }
165        if x_len < x_numel {
166            return Err(Error::BufferTooSmall {
167                needed: x_numel as usize,
168                got: x_len as usize,
169            });
170        }
171        Ok(())
172    }
173
174    /// Workspace size in bytes. Always 0 for the naive trailblazer.
175    #[inline]
176    pub fn workspace_size(&self) -> usize {
177        0
178    }
179    /// Identity of the kernel this plan picked.
180    #[inline]
181    pub fn sku(&self) -> KernelSku {
182        self.sku
183    }
184    /// Numerical guarantees.
185    #[inline]
186    pub fn precision_guarantee(&self) -> PrecisionGuarantee {
187        self.sku.precision_guarantee
188    }
189
190    /// Launch.
191    pub fn run(
192        &self,
193        stream: &Stream,
194        _workspace: Workspace<'_>,
195        args: BoolReduceArgs<'_, T, N>,
196    ) -> Result<()> {
197        self.can_implement(&args)?;
198        let output_numel = args.y.numel();
199        if output_numel == 0 {
200            return Ok(());
201        }
202        let x_ptr = args.x.data.as_raw().0 as *const c_void;
203        let y_ptr = args.y.data.as_raw().0 as *mut c_void;
204        let stream_ptr = stream.as_raw() as *mut c_void;
205
206        let output_shape = self.desc.output_shape();
207        let stride_x = args.x.stride;
208        let stride_y = args.y.stride;
209        let rank = N as i32;
210        let reduce_axis = self.desc.reduce_axis as i32;
211        let reduce_extent = self.desc.input_shape[self.desc.reduce_axis as usize];
212        let reduce_stride_x = args.x.stride[self.desc.reduce_axis as usize];
213
214        macro_rules! dispatch {
215            ($sym:ident) => {{
216                unsafe {
217                    baracuda_kernels_sys::$sym(
218                        output_numel,
219                        rank,
220                        output_shape.as_ptr(),
221                        stride_x.as_ptr(),
222                        stride_y.as_ptr(),
223                        reduce_axis,
224                        reduce_extent,
225                        reduce_stride_x,
226                        x_ptr,
227                        y_ptr,
228                        core::ptr::null_mut(),
229                        0,
230                        stream_ptr,
231                    )
232                }
233            }};
234        }
235
236        let status = match (self.desc.kind, T::KIND) {
237            // Any
238            (ReduceKind::Any, ElementKind::F32) => dispatch!(baracuda_kernels_reduce_any_f32_run),
239            (ReduceKind::Any, ElementKind::F16) => dispatch!(baracuda_kernels_reduce_any_f16_run),
240            (ReduceKind::Any, ElementKind::Bf16) => dispatch!(baracuda_kernels_reduce_any_bf16_run),
241            (ReduceKind::Any, ElementKind::F64) => dispatch!(baracuda_kernels_reduce_any_f64_run),
242            (ReduceKind::Any, ElementKind::I32) => dispatch!(baracuda_kernels_reduce_any_i32_run),
243            (ReduceKind::Any, ElementKind::I64) => dispatch!(baracuda_kernels_reduce_any_i64_run),
244            (ReduceKind::Any, ElementKind::Bool) => dispatch!(baracuda_kernels_reduce_any_bool_run),
245            // All
246            (ReduceKind::All, ElementKind::F32) => dispatch!(baracuda_kernels_reduce_all_f32_run),
247            (ReduceKind::All, ElementKind::F16) => dispatch!(baracuda_kernels_reduce_all_f16_run),
248            (ReduceKind::All, ElementKind::Bf16) => dispatch!(baracuda_kernels_reduce_all_bf16_run),
249            (ReduceKind::All, ElementKind::F64) => dispatch!(baracuda_kernels_reduce_all_f64_run),
250            (ReduceKind::All, ElementKind::I32) => dispatch!(baracuda_kernels_reduce_all_i32_run),
251            (ReduceKind::All, ElementKind::I64) => dispatch!(baracuda_kernels_reduce_all_i64_run),
252            (ReduceKind::All, ElementKind::Bool) => dispatch!(baracuda_kernels_reduce_all_bool_run),
253            _ => {
254                return Err(Error::Unsupported(
255                    "baracuda-kernels::BoolReducePlan::run: only `{Any, All} × \
256                     {f32, f16, bf16, f64, i32, i64, Bool}` wired",
257                ));
258            }
259        };
260        map_status(status)
261    }
262}
263
264fn map_status(code: i32) -> Result<()> {
265    match code {
266        0 => Ok(()),
267        1 => Err(Error::MisalignedOperand),
268        2 => Err(Error::InvalidProblem(
269            "baracuda-kernels-sys reported invalid problem",
270        )),
271        3 => Err(Error::Unsupported(
272            "baracuda-kernels-sys reported unsupported configuration",
273        )),
274        4 => Err(Error::WorkspaceTooSmall { needed: 0, got: 0 }),
275        n => Err(Error::CutlassInternal(n)),
276    }
277}
baracuda_kernels/reduce/bool_axis.rs

baracuda_kernels/reduce/
bool_axis.rs
