primal-barracuda 0.4.0

barraCuda — BARrier-free Rust Abstracted Cross-platform Unified Dimensional Algebra. Vendor-agnostic GPU compute via WGSL.
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
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Buffer pool for zero-overhead tensor operations
//!
//! Buffers returned to the pool are held in a pending queue until a
//! non-blocking `device.poll(MaintainBase::Poll)` confirms that all
//! previously submitted GPU work has progressed. This prevents the
//! "drop-before-completion" race (S-13) where a recycled buffer could
//! be handed out while the GPU is still reading/writing its contents.

use crate::error::{BarracudaError, Result};
use std::collections::HashMap;
use std::ops::Deref;
use std::sync::atomic::{AtomicU32, AtomicUsize, Ordering};
use std::sync::{Arc, Mutex, RwLock, Weak};

/// A buffer that automatically returns to its pool when dropped.
///
/// The buffer enters a pending queue on drop rather than being
/// immediately available for reuse — see [`BufferPool::drain_pending`].
pub struct PooledBuffer {
    buffer: Option<wgpu::Buffer>,
    pool: Weak<BufferPoolInner>,
    bucket: usize,
}

impl PooledBuffer {
    pub(crate) fn new(buffer: wgpu::Buffer, pool: Weak<BufferPoolInner>, bucket: usize) -> Self {
        Self {
            buffer: Some(buffer),
            pool,
            bucket,
        }
    }

    /// Get the underlying wgpu buffer.
    ///
    /// # Panics
    /// Never panics: buffer is `None` only after [`into_buffer`](Self::into_buffer) (consumes self)
    /// or during `Drop`; this method is not used in those states.
    #[must_use]
    #[expect(
        clippy::expect_used,
        reason = "Option::Some invariant: buffer is taken only by into_buffer/Drop, which consume self"
    )]
    pub fn buffer(&self) -> &wgpu::Buffer {
        // Invariant: buffer is Some from construction until into_buffer/Drop.
        self.buffer
            .as_ref()
            .expect("PooledBuffer::buffer after into_buffer (unreachable by ownership)")
    }

    /// Get the buffer size in bytes.
    #[must_use]
    pub fn size(&self) -> u64 {
        self.buffer().size()
    }

    /// Convert to a regular `wgpu::Buffer` (removes from pool management).
    ///
    /// # Panics
    ///
    /// Unreachable by construction: `into_buffer` consumes `self` so it
    /// cannot be called twice.
    #[must_use]
    #[expect(
        clippy::expect_used,
        reason = "into_buffer consumes self; Option::None here means double-call (unreachable)"
    )]
    pub fn into_buffer(mut self) -> wgpu::Buffer {
        // Invariant: buffer is Some; into_buffer consumes self so only called once.
        self.buffer
            .take()
            .expect("PooledBuffer::into_buffer called twice (unreachable by ownership)")
    }
}

impl Deref for PooledBuffer {
    type Target = wgpu::Buffer;

    #[expect(
        clippy::expect_used,
        reason = "Deref requires &T; buffer is Some until into_buffer/Drop consume self"
    )]
    fn deref(&self) -> &Self::Target {
        self.buffer
            .as_ref()
            .expect("PooledBuffer::deref after into_buffer (unreachable by ownership)")
    }
}

impl Drop for PooledBuffer {
    fn drop(&mut self) {
        if let Some(buffer) = self.buffer.take()
            && let Some(pool) = self.pool.upgrade()
        {
            pool.defer_return(buffer, self.bucket);
        }
    }
}

/// Inner pool structure (separate to allow Weak references)
pub(crate) struct BufferPoolInner {
    pub pools: RwLock<HashMap<usize, Vec<wgpu::Buffer>>>,
    pub device: wgpu::Device,
    /// Serializes poll against submit (shared with `WgpuDevice::gpu_lock`).
    poll_lock: Arc<Mutex<()>>,
    /// Counts live `GuardedEncoders` — poll skips when > 0.
    active_encoders: Arc<AtomicU32>,
    pub allocations: AtomicUsize,
    pub reuses: AtomicUsize,
    pub solver_buffers: RwLock<HashMap<String, SolverBufferSet>>,
    pending: Mutex<Vec<(wgpu::Buffer, usize)>>,
}

impl BufferPoolInner {
    fn return_buffer(&self, buffer: wgpu::Buffer, bucket: usize) {
        self.pools
            .write()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .entry(bucket)
            .or_default()
            .push(buffer);
        self.reuses.fetch_add(1, Ordering::Relaxed);
    }

    /// Place a buffer into the pending queue instead of making it
    /// immediately available. The buffer becomes reusable only after
    /// [`BufferPool::drain_pending`] calls `device.poll`.
    fn defer_return(&self, buffer: wgpu::Buffer, bucket: usize) {
        self.pending
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .push((buffer, bucket));
    }
}

/// Descriptor for creating a pinned solver buffer.
#[derive(Debug, Clone)]
pub struct BufferDescriptor {
    /// Buffer size in bytes.
    pub size: u64,
    /// wgpu buffer usage flags.
    pub usage: wgpu::BufferUsages,
    /// Optional debug label.
    pub label: Option<String>,
}

impl BufferDescriptor {
    /// Create a descriptor with default storage + copy usage.
    #[must_use]
    pub fn new(size: u64) -> Self {
        Self {
            size,
            usage: wgpu::BufferUsages::STORAGE
                | wgpu::BufferUsages::COPY_SRC
                | wgpu::BufferUsages::COPY_DST,
            label: None,
        }
    }

    /// Descriptor for an f64 array of given count.
    #[must_use]
    pub fn f64_array(count: usize) -> Self {
        Self::new((count * std::mem::size_of::<f64>()) as u64)
    }

    /// Descriptor for an f32 array of given count.
    #[must_use]
    pub fn f32_array(count: usize) -> Self {
        Self::new((count * std::mem::size_of::<f32>()) as u64)
    }

    /// Set the buffer label.
    #[must_use]
    pub fn with_label(mut self, label: impl Into<String>) -> Self {
        self.label = Some(label.into());
        self
    }

    /// Set buffer usage flags.
    #[must_use]
    pub fn with_usage(mut self, usage: wgpu::BufferUsages) -> Self {
        self.usage = usage;
        self
    }
}

/// A set of buffers pinned for the lifetime of a solver.
#[derive(Debug)]
pub struct SolverBufferSet {
    solver_id: String,
    buffers: HashMap<String, Arc<wgpu::Buffer>>,
}

impl SolverBufferSet {
    /// Get a buffer by name.
    #[must_use]
    pub fn get(&self, name: &str) -> Option<&wgpu::Buffer> {
        self.buffers.get(name).map(std::convert::AsRef::as_ref)
    }

    /// Get a buffer as Arc by name.
    #[must_use]
    pub fn get_arc(&self, name: &str) -> Option<Arc<wgpu::Buffer>> {
        self.buffers.get(name).cloned()
    }

    /// Solver identifier.
    #[must_use]
    pub fn solver_id(&self) -> &str {
        &self.solver_id
    }

    /// Iterator over buffer names.
    pub fn buffer_names(&self) -> impl Iterator<Item = &str> {
        self.buffers.keys().map(std::string::String::as_str)
    }

    /// Number of buffers in the set.
    #[must_use]
    pub fn len(&self) -> usize {
        self.buffers.len()
    }

    /// Returns true if the set has no buffers.
    #[must_use]
    pub fn is_empty(&self) -> bool {
        self.buffers.is_empty()
    }
}

/// Memory pool for buffer reuse
pub struct BufferPool {
    inner: Arc<BufferPoolInner>,
}

impl BufferPool {
    /// Create a new buffer pool with shared locks.
    pub fn new(
        device: wgpu::Device,
        poll_lock: Arc<Mutex<()>>,
        active_encoders: Arc<AtomicU32>,
    ) -> Self {
        Self {
            inner: Arc::new(BufferPoolInner {
                pools: RwLock::new(HashMap::new()),
                device,
                poll_lock,
                active_encoders,
                allocations: AtomicUsize::new(0),
                reuses: AtomicUsize::new(0),
                solver_buffers: RwLock::new(HashMap::new()),
                pending: Mutex::new(Vec::new()),
            }),
        }
    }

    /// Create a pool with its own independent locks (for tests/standalone use).
    #[must_use]
    pub fn new_standalone(device: wgpu::Device) -> Self {
        Self::new(
            device,
            Arc::new(Mutex::new(())),
            Arc::new(AtomicU32::new(0)),
        )
    }

    fn bucket_size(size: usize) -> usize {
        let min_size = 256;
        let size = size.max(min_size);
        size.next_power_of_two()
    }

    /// Try to move pending buffers back into the available pool.
    /// Uses `try_write()` on the `poll_lock` so it **never blocks encoding**.
    /// If encoding is in progress (read locks held), the drain is skipped
    /// entirely — buffers stay pending until the next drain attempt.
    /// Uses `MaintainBase::Poll` (non-blocking) to check for GPU completion
    /// without stalling. Buffers whose GPU work hasn't finished remain pending.
    /// # Panics
    /// Panics if the pending lock is poisoned.
    pub fn drain_pending(&self) {
        let pending_count = self
            .inner
            .pending
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .len();
        if pending_count == 0 {
            return;
        }
        if self.inner.active_encoders.load(Ordering::Acquire) > 0 {
            return;
        }
        let _guard = match self.inner.poll_lock.try_lock() {
            Ok(g) => g,
            Err(std::sync::TryLockError::WouldBlock) => return,
            Err(std::sync::TryLockError::Poisoned(e)) => e.into_inner(),
        };
        let _ = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            let _ = self.inner.device.poll(wgpu::PollType::Poll);
        }));

        let drained: Vec<(wgpu::Buffer, usize)> = {
            let mut pending = self
                .inner
                .pending
                .lock()
                .unwrap_or_else(std::sync::PoisonError::into_inner);
            std::mem::take(&mut *pending)
        };

        for (buffer, bucket) in drained {
            self.inner.return_buffer(buffer, bucket);
        }
    }

    /// Acquire a pooled buffer (returns to pool on drop).
    /// # Panics
    /// Panics if the pools lock is poisoned.
    #[must_use]
    pub fn acquire_pooled(&self, size_bytes: usize) -> PooledBuffer {
        self.drain_pending();

        let bucket = Self::bucket_size(size_bytes);
        let buffer = self
            .inner
            .pools
            .write()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .get_mut(&bucket)
            .and_then(std::vec::Vec::pop)
            .unwrap_or_else(|| self.allocate_new(bucket));
        PooledBuffer::new(buffer, Arc::downgrade(&self.inner), bucket)
    }

    /// Acquire a buffer (does not return to pool; call release explicitly).
    /// # Panics
    /// Panics if the pools lock is poisoned.
    #[must_use]
    pub fn acquire(&self, size_bytes: usize) -> wgpu::Buffer {
        self.drain_pending();

        let bucket = Self::bucket_size(size_bytes);
        let recycled = self
            .inner
            .pools
            .write()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .get_mut(&bucket)
            .and_then(std::vec::Vec::pop);
        if let Some(buf) = recycled {
            self.inner.reuses.fetch_add(1, Ordering::Relaxed);
            return buf;
        }
        self.allocate_new(bucket)
    }

    fn allocate_new(&self, bucket: usize) -> wgpu::Buffer {
        self.inner.allocations.fetch_add(1, Ordering::Relaxed);
        self.inner.active_encoders.fetch_add(1, Ordering::Acquire);
        let buf = self.inner.device.create_buffer(&wgpu::BufferDescriptor {
            label: Some("Pooled Buffer"),
            size: bucket as u64,
            usage: wgpu::BufferUsages::STORAGE
                | wgpu::BufferUsages::COPY_SRC
                | wgpu::BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });
        self.inner.active_encoders.fetch_sub(1, Ordering::Release);
        buf
    }

    /// Return a buffer to the pool for reuse.
    /// # Panics
    /// Panics if the pools lock is poisoned.
    pub fn release(&self, buffer: wgpu::Buffer) {
        let bucket = Self::bucket_size(buffer.size() as usize);
        self.inner
            .pools
            .write()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .entry(bucket)
            .or_default()
            .push(buffer);
    }

    /// Returns (allocations, reuses).
    #[must_use]
    pub fn stats(&self) -> (usize, usize) {
        (
            self.inner.allocations.load(Ordering::Relaxed),
            self.inner.reuses.load(Ordering::Relaxed),
        )
    }

    /// Pin buffers for a solver; returns error if solver already has buffers.
    /// # Panics
    /// Panics if the `solver_buffers` lock is poisoned.
    /// # Errors
    /// Returns [`Err`] if the solver already has pinned buffers (call
    /// `release_solver_buffers` first), or if buffer allocation fails.
    pub fn pin_solver_buffers(
        &self,
        solver_id: &str,
        buffers: &[(&str, BufferDescriptor)],
    ) -> Result<SolverBufferSet> {
        let mut solver_buffers = self
            .inner
            .solver_buffers
            .write()
            .unwrap_or_else(std::sync::PoisonError::into_inner);

        if solver_buffers.contains_key(solver_id) {
            return Err(BarracudaError::InvalidInput {
                message: format!(
                    "Solver '{solver_id}' already has pinned buffers. Call release_solver_buffers() first."
                ),
            });
        }

        let mut buffer_map = HashMap::new();
        for (name, desc) in buffers {
            let label = desc
                .label
                .clone()
                .unwrap_or_else(|| format!("{solver_id}:{name}"));
            let buffer = self.inner.device.create_buffer(&wgpu::BufferDescriptor {
                label: Some(&label),
                size: desc.size,
                usage: desc.usage,
                mapped_at_creation: false,
            });
            self.inner.allocations.fetch_add(1, Ordering::Relaxed);
            buffer_map.insert(name.to_string(), Arc::new(buffer));
        }

        let buffer_set = SolverBufferSet {
            solver_id: solver_id.to_string(),
            buffers: buffer_map.clone(),
        };
        solver_buffers.insert(solver_id.to_string(), buffer_set);

        Ok(SolverBufferSet {
            solver_id: solver_id.to_string(),
            buffers: buffer_map,
        })
    }

    /// Release pinned buffers for a solver. Returns true if buffers were released.
    /// # Panics
    /// Panics if the `solver_buffers` lock is poisoned.
    #[must_use]
    pub fn release_solver_buffers(&self, solver_id: &str) -> bool {
        let mut solver_buffers = self
            .inner
            .solver_buffers
            .write()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        solver_buffers.remove(solver_id).is_some()
    }

    /// Get pinned buffers for a solver (cloned).
    /// # Panics
    /// Panics if the `solver_buffers` lock is poisoned.
    #[must_use]
    pub fn get_solver_buffers(&self, solver_id: &str) -> Option<SolverBufferSet> {
        let solver_buffers = self
            .inner
            .solver_buffers
            .read()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        solver_buffers.get(solver_id).map(|set| SolverBufferSet {
            solver_id: set.solver_id.clone(),
            buffers: set.buffers.clone(),
        })
    }

    /// Returns true if solver has pinned buffers.
    /// # Panics
    /// Panics if the `solver_buffers` lock is poisoned.
    #[must_use]
    pub fn has_solver_buffers(&self, solver_id: &str) -> bool {
        let solver_buffers = self
            .inner
            .solver_buffers
            .read()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        solver_buffers.contains_key(solver_id)
    }

    /// All solver IDs with pinned buffers.
    /// # Panics
    /// Panics if the `solver_buffers` lock is poisoned.
    #[must_use]
    pub fn solver_ids(&self) -> Vec<String> {
        let solver_buffers = self
            .inner
            .solver_buffers
            .read()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        solver_buffers.keys().cloned().collect()
    }
}