krylovscope 0.1.0

GPU-resident Krylov spectral endpoint estimation for large symmetric operators
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
#![allow(clippy::assign_op_pattern)]

//! Built-in sparse operators and validation for RobustF64Q.
//!
//! Host storage is borrowed only during validation/upload.  The returned CUDA
//! operator owns device handles, so solver-side CSR/PDSM buffers may be
//! released or reused after construction.  Expensive structural/symmetry
//! checks are intentionally host-side setup work, not Lanczos hot-path work.
use crate::cuda::{CsrF32Operator, ResidentF64Operator};
use cubecl::{client::ComputeClient, prelude::*, server::Handle};
use std::collections::HashMap;

const WG: u32 = 256;
const PDSM_WINDOW: usize = 32;

#[derive(Clone, Copy, Debug)]
pub enum HostSparseOperator<'a> {
    CsrF32 {
        row_ptr: &'a [u32],
        col_idx: &'a [u32],
        values: &'a [f32],
    },
    Pdsm32F32 {
        base_col: &'a [u32],
        bitmap: &'a [u32],
        row_value_offset: &'a [u32],
        values: &'a [f32],
    },
}

#[derive(Clone, Copy, Debug)]
pub struct ValidationPolicy {
    pub require_sorted_columns: bool,
    pub require_diagonal: bool,
    pub require_symmetric: bool,
    /// Relative tolerance used for each `(i, j)` / `(j, i)` pair.
    /// Must be finite and nonnegative when symmetry validation is enabled.
    pub symmetry_tol: f64,
}
impl Default for ValidationPolicy {
    fn default() -> Self {
        Self {
            require_sorted_columns: true,
            require_diagonal: true,
            require_symmetric: true,
            symmetry_tol: 2.0e-6,
        }
    }
}

#[derive(Clone, Debug)]
pub struct ValidationReport {
    pub kind: &'static str,
    pub n: usize,
    pub nnz: usize,
    pub sorted_columns: bool,
    pub diagonal_complete: bool,
    pub symmetric_checked: bool,
    pub max_symmetry_abs: f64,
}

fn key(row: usize, col: usize) -> u64 {
    ((row as u64) << 32) | col as u64
}

fn check_symmetry(n: usize, entries: &[(usize, usize, f32)], tol: f64) -> Result<f64, String> {
    if n > u32::MAX as usize {
        return Err("krylovscope: dimension exceeds u32 index space".into());
    }
    let mut map = HashMap::<u64, f32>::with_capacity(entries.len() * 2);
    for &(r, c, v) in entries {
        if map.insert(key(r, c), v).is_some() {
            return Err(format!("krylovscope: duplicate matrix entry at ({r},{c})"));
        }
    }
    let mut worst = 0.0_f64;
    for &(r, c, v) in entries {
        let Some(&vt) = map.get(&key(c, r)) else {
            return Err(format!(
                "krylovscope: symmetry partner missing for ({r},{c})"
            ));
        };
        let d = (v as f64 - vt as f64).abs();
        worst = worst.max(d);
        let scale = (v as f64).abs().max((vt as f64).abs());
        if d > tol * scale {
            return Err(format!(
                "krylovscope: symmetry mismatch at ({r},{c}): abs={d:.3e}"
            ));
        }
    }
    Ok(worst)
}

pub fn validate_host_operator(
    op: HostSparseOperator<'_>,
    policy: ValidationPolicy,
) -> Result<ValidationReport, String> {
    if policy.require_symmetric && (!policy.symmetry_tol.is_finite() || policy.symmetry_tol < 0.0) {
        return Err(format!(
            "krylovscope: symmetry_tol must be finite and nonnegative, got {}",
            policy.symmetry_tol
        ));
    }
    match op {
        HostSparseOperator::CsrF32 {
            row_ptr,
            col_idx,
            values,
        } => {
            if row_ptr.len() < 2 {
                return Err("krylovscope: CSR row_ptr too short".into());
            }
            if row_ptr[0] != 0 {
                return Err("krylovscope: CSR row_ptr[0] must be zero".into());
            }
            let n = row_ptr.len() - 1;
            if n > u32::MAX as usize {
                return Err("krylovscope: CSR dimension exceeds u32".into());
            }
            if col_idx.len() != values.len() || row_ptr[n] as usize != values.len() {
                return Err("krylovscope: CSR nnz/row_ptr mismatch".into());
            }
            let mut diag = vec![false; n];
            let mut entries = Vec::with_capacity(values.len());
            let mut sorted = true;
            for r in 0..n {
                let b = row_ptr[r] as usize;
                let e = row_ptr[r + 1] as usize;
                if b > e || e > values.len() {
                    return Err(format!("krylovscope: CSR bad row range at {r}"));
                }
                let mut prev = None;
                for k in b..e {
                    let c = col_idx[k] as usize;
                    if c >= n {
                        return Err(format!("krylovscope: CSR column out of range at row {r}"));
                    }
                    if let Some(p) = prev {
                        if c <= p {
                            sorted = false;
                        }
                    }
                    prev = Some(c);
                    if c == r {
                        diag[r] = true;
                    }
                    if !values[k].is_finite() {
                        return Err(format!("krylovscope: CSR nonfinite value at ({r},{c})"));
                    }
                    entries.push((r, c, values[k]));
                }
            }
            if policy.require_sorted_columns && !sorted {
                return Err("krylovscope: CSR columns must be strictly sorted per row".into());
            }
            let diag_ok = diag.iter().all(|&x| x);
            if policy.require_diagonal && !diag_ok {
                return Err("krylovscope: CSR diagonal is incomplete".into());
            }
            let max_sym = if policy.require_symmetric {
                check_symmetry(n, &entries, policy.symmetry_tol)?
            } else {
                0.0
            };
            Ok(ValidationReport {
                kind: "csr-f32",
                n,
                nnz: values.len(),
                sorted_columns: sorted,
                diagonal_complete: diag_ok,
                symmetric_checked: policy.require_symmetric,
                max_symmetry_abs: max_sym,
            })
        }
        HostSparseOperator::Pdsm32F32 {
            base_col,
            bitmap,
            row_value_offset,
            values,
        } => {
            let n = base_col.len();
            if n == 0 || bitmap.len() != n || row_value_offset.len() != n + 1 {
                return Err("krylovscope: malformed PDSM32 metadata".into());
            }
            if row_value_offset[0] != 0 || row_value_offset[n] as usize != values.len() {
                return Err("krylovscope: PDSM32 packed offset mismatch".into());
            }
            if n > u32::MAX as usize {
                return Err("krylovscope: PDSM32 dimension exceeds u32".into());
            }
            let mut diag = vec![false; n];
            let mut entries = Vec::with_capacity(values.len());
            for r in 0..n {
                let b = row_value_offset[r] as usize;
                let e = row_value_offset[r + 1] as usize;
                if b > e || e > values.len() {
                    return Err(format!("krylovscope: PDSM32 bad row offset at {r}"));
                }
                let base = base_col[r] as usize;
                let mask = bitmap[r];
                if mask.count_ones() as usize != e - b {
                    return Err(format!(
                        "krylovscope: PDSM32 bitmap/value count mismatch at row {r}"
                    ));
                }
                let mut k = b;
                for slot in 0..PDSM_WINDOW {
                    if (mask & (1u32 << slot)) != 0 {
                        let c = base
                            .checked_add(slot)
                            .ok_or("krylovscope: PDSM32 column overflow")?;
                        if c >= n {
                            return Err(format!(
                                "krylovscope: PDSM32 column out of range at row {r}"
                            ));
                        }
                        let v = values[k];
                        if !v.is_finite() {
                            return Err(format!(
                                "krylovscope: PDSM32 nonfinite value at ({r},{c})"
                            ));
                        }
                        if c == r {
                            diag[r] = true;
                        }
                        entries.push((r, c, v));
                        k += 1;
                    }
                }
                if k != e {
                    return Err(format!(
                        "krylovscope: PDSM32 packed scan mismatch at row {r}"
                    ));
                }
            }
            let diag_ok = diag.iter().all(|&x| x);
            if policy.require_diagonal && !diag_ok {
                return Err("krylovscope: PDSM32 diagonal is incomplete".into());
            }
            let max_sym = if policy.require_symmetric {
                check_symmetry(n, &entries, policy.symmetry_tol)?
            } else {
                0.0
            };
            Ok(ValidationReport {
                kind: "pdsm32-f32",
                n,
                nnz: values.len(),
                sorted_columns: true,
                diagonal_complete: diag_ok,
                symmetric_checked: policy.require_symmetric,
                max_symmetry_abs: max_sym,
            })
        }
    }
}

#[cube(launch_unchecked)]
fn pdsm32_f32_f64(
    base_col: &mut Array<u32>,
    bitmap: &mut Array<u32>,
    row_value_offset: &mut Array<u32>,
    values: &Array<f32>,
    x: &Array<f64>,
    y: &mut Array<f64>,
) {
    let row = ABSOLUTE_POS;
    if row < y.len() {
        let base = usize::cast_from(base_col[row]);
        let mask = bitmap[row];
        let mut k = usize::cast_from(row_value_offset[row]);
        let mut slot = 0usize;
        let mut bit = 1u32;
        let mut sum = 0.0_f64;
        while slot < PDSM_WINDOW {
            if (mask & bit) != 0u32 {
                sum += f64::cast_from(values[k]) * x[base + slot];
                k += 1usize;
            }
            slot += 1usize;
            bit = bit * 2u32;
        }
        y[row] = sum;
    }
}

pub struct Pdsm32F32Operator<R: Runtime> {
    base_col: Handle,
    bitmap: Handle,
    row_value_offset: Handle,
    values: Handle,
    n: usize,
    nnz: usize,
    _runtime: core::marker::PhantomData<R>,
}
impl<R: Runtime> Pdsm32F32Operator<R> {
    pub(crate) fn upload(
        client: &ComputeClient<R>,
        base_col: &[u32],
        bitmap: &[u32],
        row_value_offset: &[u32],
        values: &[f32],
    ) -> Self {
        Self {
            base_col: client.create_from_slice(bytemuck::cast_slice(base_col)),
            bitmap: client.create_from_slice(bytemuck::cast_slice(bitmap)),
            row_value_offset: client.create_from_slice(bytemuck::cast_slice(row_value_offset)),
            values: client.create_from_slice(bytemuck::cast_slice(values)),
            n: base_col.len(),
            nnz: values.len(),
            _runtime: core::marker::PhantomData,
        }
    }
    pub fn nnz(&self) -> usize {
        self.nnz
    }
}
impl<R: Runtime> ResidentF64Operator<R> for Pdsm32F32Operator<R> {
    fn len(&self) -> usize {
        self.n
    }
    fn apply(&mut self, client: &ComputeClient<R>, x: &Handle, y: &Handle) {
        unsafe {
            pdsm32_f32_f64::launch_unchecked::<R>(
                client,
                CubeCount::Static((self.n as u32).div_ceil(WG), 1, 1),
                CubeDim::new_1d(WG),
                ArrayArg::from_raw_parts(self.base_col.clone(), self.n),
                ArrayArg::from_raw_parts(self.bitmap.clone(), self.n),
                ArrayArg::from_raw_parts(self.row_value_offset.clone(), self.n + 1),
                ArrayArg::from_raw_parts(self.values.clone(), self.nnz),
                ArrayArg::from_raw_parts(x.clone(), self.n),
                ArrayArg::from_raw_parts(y.clone(), self.n),
            );
        }
    }
    fn name(&self) -> &'static str {
        "pdsm32-f32-to-f64"
    }
}

pub enum UploadedResidentOperator<R: Runtime> {
    Csr(CsrF32Operator<R>),
    Pdsm32(Pdsm32F32Operator<R>),
}
impl<R: Runtime> ResidentF64Operator<R> for UploadedResidentOperator<R> {
    fn len(&self) -> usize {
        match self {
            Self::Csr(x) => x.len(),
            Self::Pdsm32(x) => x.len(),
        }
    }
    fn apply(&mut self, client: &ComputeClient<R>, x: &Handle, y: &Handle) {
        match self {
            Self::Csr(op) => op.apply(client, x, y),
            Self::Pdsm32(op) => op.apply(client, x, y),
        }
    }
    fn dispatches_per_apply(&self) -> usize {
        1
    }
    fn name(&self) -> &'static str {
        match self {
            Self::Csr(_) => "csr-f32-to-f64",
            Self::Pdsm32(_) => "pdsm32-f32-to-f64",
        }
    }
}

pub fn validate_and_upload<R: Runtime>(
    client: &ComputeClient<R>,
    op: HostSparseOperator<'_>,
    policy: ValidationPolicy,
) -> Result<(UploadedResidentOperator<R>, ValidationReport), String> {
    let report = validate_host_operator(op, policy)?;
    let uploaded = match op {
        HostSparseOperator::CsrF32 {
            row_ptr,
            col_idx,
            values,
        } => UploadedResidentOperator::Csr(CsrF32Operator::new(client, row_ptr, col_idx, values)?),
        HostSparseOperator::Pdsm32F32 {
            base_col,
            bitmap,
            row_value_offset,
            values,
        } => UploadedResidentOperator::Pdsm32(Pdsm32F32Operator::upload(
            client,
            base_col,
            bitmap,
            row_value_offset,
            values,
        )),
    };
    Ok((uploaded, report))
}

#[cfg(test)]
mod tests {
    use super::*;

    fn asymmetric_csr<'a>(
        row_ptr: &'a [u32],
        col_idx: &'a [u32],
        values: &'a [f32],
    ) -> HostSparseOperator<'a> {
        HostSparseOperator::CsrF32 {
            row_ptr,
            col_idx,
            values,
        }
    }

    #[test]
    fn nonfinite_symmetry_tolerance_is_rejected() {
        let row_ptr = [0, 2, 4];
        let col_idx = [0, 1, 0, 1];
        let values = [1.0, 2.0, 3.0, 1.0];
        for tol in [f64::NAN, f64::INFINITY, f64::NEG_INFINITY] {
            let policy = ValidationPolicy {
                symmetry_tol: tol,
                ..ValidationPolicy::default()
            };
            let error = validate_host_operator(asymmetric_csr(&row_ptr, &col_idx, &values), policy)
                .unwrap_err();
            assert!(error.contains("symmetry_tol"));
        }
    }

    #[test]
    fn negative_symmetry_tolerance_is_rejected() {
        let row_ptr = [0, 2, 4];
        let col_idx = [0, 1, 0, 1];
        let values = [1.0, 2.0, 3.0, 1.0];
        let policy = ValidationPolicy {
            symmetry_tol: -1.0,
            ..ValidationPolicy::default()
        };
        let error = validate_host_operator(asymmetric_csr(&row_ptr, &col_idx, &values), policy)
            .unwrap_err();
        assert!(error.contains("symmetry_tol"));
    }

    #[test]
    fn symmetry_tolerance_is_relative_at_small_scale() {
        let row_ptr = [0, 2, 4];
        let col_idx = [0, 1, 0, 1];
        let values = [1.0e-8, 2.0e-8, 3.0e-8, 1.0e-8];
        let error = validate_host_operator(
            asymmetric_csr(&row_ptr, &col_idx, &values),
            ValidationPolicy::default(),
        )
        .unwrap_err();
        assert!(error.contains("symmetry mismatch"));
    }
}