gpu-alloc 0.4.5

Implementation agnostic memory allocator for Vulkan like APIs
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
use {
    crate::{
        align_down, align_up,
        error::AllocationError,
        heap::Heap,
        util::{arc_unwrap, is_arc_unique},
        MemoryBounds,
    },
    alloc::sync::Arc,
    core::{cmp::Ordering, ptr::NonNull},
    gpu_alloc_types::{AllocationFlags, DeviceMapError, MemoryDevice, MemoryPropertyFlags},
};

unsafe fn opt_ptr_add(ptr: Option<NonNull<u8>>, size: u64) -> Option<NonNull<u8>> {
    ptr.map(|ptr| {
        // Size is within memory region started at `ptr`.
        // size is within `chunk_size` that fits `isize`.
        NonNull::new_unchecked(ptr.as_ptr().offset(size as isize))
    })
}

#[derive(Debug)]
pub(super) struct FreeList<M> {
    array: Vec<FreeListRegion<M>>,
    total: u64,
    counter: u64,
}

impl<M> FreeList<M> {
    pub fn new() -> Self {
        FreeList {
            array: Vec::new(),
            total: 0,
            counter: 0,
        }
    }

    pub fn get_block_from_new_memory(
        &mut self,
        memory: Arc<M>,
        memory_size: u64,
        ptr: Option<NonNull<u8>>,
        align_mask: u64,
        size: u64,
    ) -> FreeListBlock<M> {
        debug_assert!(size <= memory_size);

        self.counter += 1;
        self.array.push(FreeListRegion {
            memory,
            ptr,
            chunk: self.counter,
            start: 0,
            end: memory_size,
        });
        self.total += memory_size;
        self.get_block_at(self.array.len() - 1, align_mask, size)
    }

    pub fn get_block(&mut self, align_mask: u64, size: u64) -> Option<FreeListBlock<M>> {
        match &self.array[..] {
            [] => None,
            [rest @ .., last] => {
                let (index, _) = std::iter::once((rest.len(), last))
                    .chain(rest.iter().enumerate())
                    .find(|(_, region)| match region.end.checked_sub(size) {
                        Some(start) => {
                            let aligned_start = align_down(start, align_mask);
                            aligned_start >= region.start
                        }
                        None => false,
                    })?;

                Some(self.get_block_at(index, align_mask, size))
            }
        }
    }

    fn get_block_at(&mut self, index: usize, align_mask: u64, size: u64) -> FreeListBlock<M> {
        let region = &mut self.array[index];

        let start = region.end - size;
        let aligned_start = align_down(start, align_mask);

        if aligned_start > region.start {
            let block = FreeListBlock {
                offset: aligned_start,
                size: region.end - aligned_start,
                chunk: region.chunk,
                ptr: unsafe { opt_ptr_add(region.ptr, aligned_start - region.start) },
                memory: region.memory.clone(),
            };

            region.end = aligned_start;
            self.total -= block.size;

            block
        } else {
            debug_assert_eq!(aligned_start, region.start);
            let region = self.array.remove(index);
            self.total -= region.end - region.start;
            region.into_block()
        }
    }

    pub fn insert_block(&mut self, block: FreeListBlock<M>) {
        let block_size = block.size;

        match self.array.binary_search_by(|b| b.cmp(&block)) {
            Ok(_) => {
                panic!("Overlapping block found in free list");
            }
            Err(index) if self.array.len() > index => match &mut self.array[..=index] {
                [] => unreachable!(),
                [next] => {
                    debug_assert!(!next.is_suffix_block(&block));

                    if next.is_prefix_block(&block) {
                        next.merge_prefix_block(block);
                    } else {
                        self.array.insert(0, FreeListRegion::from_block(block));
                    }
                }
                [.., prev, next] => {
                    debug_assert!(!prev.is_prefix_block(&block));
                    debug_assert!(!next.is_suffix_block(&block));

                    if next.is_prefix_block(&block) {
                        next.merge_prefix_block(block);

                        if prev.consecutive(&next) {
                            let next = self.array.remove(index);
                            let prev = &mut self.array[index - 1];
                            prev.merge(next);
                        }
                    } else if prev.is_suffix_block(&block) {
                        prev.merge_suffix_block(block);
                    } else {
                        self.array.insert(index, FreeListRegion::from_block(block));
                    }
                }
            },
            Err(_) => {
                self.array.push(FreeListRegion::from_block(block));
            }
        }
        self.total += block_size;
    }

    pub fn drain(&mut self, dealloc_threshold: u64) -> Option<impl Iterator<Item = M> + '_> {
        if self.total >= dealloc_threshold {
            // Time to deallocate

            // This code was adapted from `Vec::retain`
            let len = self.array.len();
            let mut del = 0;
            {
                let regions = &mut self.array[..];

                for i in 0..len {
                    if self.total >= dealloc_threshold && is_arc_unique(&mut regions[i].memory) {
                        del += 1;
                    } else if del > 0 {
                        regions.swap(i - del, i);
                    }
                }
            }

            if del > 0 {
                let total = &mut self.total;
                return Some(self.array.drain(len - del..).map(move |region| {
                    debug_assert_eq!(region.start, 0);
                    *total -= region.end;
                    unsafe { arc_unwrap(region.memory) }
                }));
            }
        }
        None
    }
}

#[derive(Debug)]
struct FreeListRegion<M> {
    memory: Arc<M>,
    ptr: Option<NonNull<u8>>,
    chunk: u64,
    start: u64,
    end: u64,
}

unsafe impl<M> Sync for FreeListRegion<M> where M: Sync {}
unsafe impl<M> Send for FreeListRegion<M> where M: Send {}

impl<M> FreeListRegion<M> {
    pub fn cmp(&self, block: &FreeListBlock<M>) -> Ordering {
        debug_assert_eq!(
            Arc::ptr_eq(&self.memory, &block.memory),
            self.chunk == block.chunk
        );

        if self.chunk == block.chunk {
            debug_assert_eq!(
                Ord::cmp(&self.start, &block.offset),
                Ord::cmp(&self.end, &(block.offset + block.size)),
                "Free region {{ start: {}, end: {} }} overlaps with block {{ offset: {}, size: {} }}",
                self.start,
                self.end,
                block.offset,
                block.size,
            );
        }

        Ord::cmp(&self.chunk, &block.chunk).then(Ord::cmp(&self.start, &block.offset))
    }

    fn from_block(block: FreeListBlock<M>) -> Self {
        FreeListRegion {
            memory: block.memory,
            chunk: block.chunk,
            ptr: block.ptr,
            start: block.offset,
            end: block.offset + block.size,
        }
    }

    fn into_block(self) -> FreeListBlock<M> {
        FreeListBlock {
            memory: self.memory,
            chunk: self.chunk,
            ptr: self.ptr,
            offset: self.start,
            size: self.end - self.start,
        }
    }

    fn consecutive(&self, other: &Self) -> bool {
        if self.chunk != other.chunk {
            return false;
        }

        debug_assert!(Arc::ptr_eq(&self.memory, &other.memory));

        debug_assert_eq!(
            Ord::cmp(&self.start, &other.start),
            Ord::cmp(&self.end, &other.end)
        );

        self.end == other.start
    }

    fn merge(&mut self, next: FreeListRegion<M>) {
        debug_assert!(self.consecutive(&next));
        self.end = next.end;
    }

    fn is_prefix_block(&self, block: &FreeListBlock<M>) -> bool {
        if self.chunk != block.chunk {
            return false;
        }

        debug_assert!(Arc::ptr_eq(&self.memory, &block.memory));

        debug_assert_eq!(
            Ord::cmp(&self.start, &block.offset),
            Ord::cmp(&self.end, &(block.offset + block.size))
        );

        self.start == (block.offset + block.size)
    }

    fn merge_prefix_block(&mut self, block: FreeListBlock<M>) {
        debug_assert!(self.is_prefix_block(&block));
        self.start = block.offset;
        self.ptr = block.ptr;
    }

    fn is_suffix_block(&self, block: &FreeListBlock<M>) -> bool {
        if self.chunk != block.chunk {
            return false;
        }

        debug_assert!(Arc::ptr_eq(&self.memory, &block.memory));

        debug_assert_eq!(
            Ord::cmp(&self.start, &block.offset),
            Ord::cmp(&self.end, &(block.offset + block.size))
        );

        self.end == block.offset
    }

    fn merge_suffix_block(&mut self, block: FreeListBlock<M>) {
        debug_assert!(self.is_suffix_block(&block));
        self.end += block.size;
    }
}

#[derive(Debug)]
pub struct FreeListBlock<M> {
    pub memory: Arc<M>,
    pub ptr: Option<NonNull<u8>>,
    pub chunk: u64,
    pub offset: u64,
    pub size: u64,
}

unsafe impl<M> Sync for FreeListBlock<M> where M: Sync {}
unsafe impl<M> Send for FreeListBlock<M> where M: Send {}

#[derive(Debug)]
pub(crate) struct FreeListAllocator<M> {
    freelist: FreeList<M>,
    dealloc_threshold: u64,
    chunk_size: u64,
    memory_type: u32,
    props: MemoryPropertyFlags,
    atom_mask: u64,
}

impl<M> FreeListAllocator<M>
where
    M: MemoryBounds + 'static,
{
    pub fn new(
        chunk_size: u64,
        dealloc_threshold: u64,
        memory_type: u32,
        props: MemoryPropertyFlags,
        atom_mask: u64,
    ) -> Self {
        debug_assert_eq!(align_down(chunk_size, atom_mask), chunk_size);

        let chunk_size = min(chunk_size, isize::max_value());

        FreeListAllocator {
            freelist: FreeList::new(),
            chunk_size,
            dealloc_threshold,
            memory_type,
            props,
            atom_mask,
        }
    }

    #[cfg_attr(feature = "tracing", tracing::instrument(skip(self, device)))]
    pub unsafe fn alloc(
        &mut self,
        device: &impl MemoryDevice<M>,
        size: u64,
        align_mask: u64,
        flags: AllocationFlags,
        heap: &mut Heap,
        allocations_remains: &mut u32,
    ) -> Result<FreeListBlock<M>, AllocationError> {
        debug_assert!(
            self.chunk_size >= size,
            "GpuAllocator must not request allocations equal or greater to chunks size"
        );

        let size = align_up(size, self.atom_mask).expect(
            "Any value not greater than chunk size (which is aligned) has to fit for alignment",
        );

        let align_mask = align_mask | self.atom_mask;
        let host_visible = self.host_visible();

        if let Some(block) = self.freelist.get_block(align_mask, size) {
            return Ok(block);
        }

        // New allocation is required.
        if *allocations_remains == 0 {
            return Err(AllocationError::TooManyObjects);
        }
        let mut memory = device.allocate_memory(self.chunk_size, self.memory_type, flags)?;
        *allocations_remains -= 1;
        heap.alloc(self.chunk_size);

        // Map host visible allocations
        let ptr = if host_visible {
            match device.map_memory(&mut memory, 0, self.chunk_size) {
                Ok(ptr) => Some(ptr),
                Err(DeviceMapError::MapFailed) => {
                    #[cfg(feature = "tracing")]
                    tracing::error!("Failed to map host-visible memory in linear allocator");
                    device.deallocate_memory(memory);
                    *allocations_remains += 1;
                    heap.dealloc(self.chunk_size);

                    return Err(AllocationError::OutOfHostMemory);
                }
                Err(DeviceMapError::OutOfDeviceMemory) => {
                    return Err(AllocationError::OutOfDeviceMemory);
                }
                Err(DeviceMapError::OutOfHostMemory) => {
                    return Err(AllocationError::OutOfHostMemory);
                }
            }
        } else {
            None
        };

        let memory = Arc::new(memory);
        let block =
            self.freelist
                .get_block_from_new_memory(memory, self.chunk_size, ptr, align_mask, size);

        Ok(block)
    }

    #[cfg_attr(feature = "tracing", tracing::instrument(skip(self, device)))]
    pub unsafe fn dealloc(
        &mut self,
        device: &impl MemoryDevice<M>,
        block: FreeListBlock<M>,
        heap: &mut Heap,
        allocations_remains: &mut u32,
    ) {
        debug_assert!(block.size < self.chunk_size);
        debug_assert_ne!(block.size, 0);
        self.freelist.insert_block(block);

        if let Some(memory) = self.freelist.drain(self.dealloc_threshold) {
            let chunk_size = self.chunk_size;
            memory.for_each(|m| {
                device.deallocate_memory(m);
                *allocations_remains += 1;
                heap.dealloc(chunk_size);
            });
        }
    }

    /// Deallocates leftover memory objects.
    /// Should be used before dropping.
    ///
    /// # Safety
    ///
    /// * `device` must be one with `DeviceProperties` that were provided to create this `GpuAllocator` instance
    /// * Same `device` instance must be used for all interactions with one `GpuAllocator` instance
    ///   and memory blocks allocated from it
    #[cfg_attr(feature = "tracing", tracing::instrument(skip(self, device)))]
    pub unsafe fn cleanup(&mut self, device: &impl MemoryDevice<M>) {
        if let Some(memory) = self.freelist.drain(0) {
            memory.for_each(|m| device.deallocate_memory(m));
        }
    }

    fn host_visible(&self) -> bool {
        self.props.contains(MemoryPropertyFlags::HOST_VISIBLE)
    }
}

fn min<L, R>(l: L, r: R) -> L
where
    R: core::convert::TryInto<L>,
    L: Ord,
{
    match r.try_into() {
        Ok(r) => core::cmp::min(l, r),
        Err(_) => l,
    }
}