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
//! GPU Memory Pool for efficient allocation/deallocation
#[cfg(feature = "gpu")]
mod gpu_pool {
use cudarc::driver::{CudaContext, CudaSlice};
use std::collections::{HashMap, VecDeque};
use std::sync::{Arc, Mutex};
/// Memory pool bucket for a specific size range
pub(super) struct MemoryBucket<T: Clone> {
/// Free buffers ready for reuse
free_buffers: VecDeque<CudaSlice<T>>,
/// Maximum number of buffers to keep
max_buffers: usize,
/// Buffer size for this bucket
buffer_size: usize,
}
impl<T: Clone + cudarc::driver::DeviceRepr> MemoryBucket<T> {
pub(super) fn new(buffer_size: usize, max_buffers: usize) -> Self {
Self {
free_buffers: VecDeque::new(),
max_buffers,
buffer_size,
}
}
/// Try to get a buffer from the free list
pub(super) fn try_alloc(&mut self) -> Option<CudaSlice<T>> {
self.free_buffers.pop_front()
}
/// Return a buffer to the free list (if not full)
pub(super) fn free(&mut self, buffer: CudaSlice<T>) {
if self.free_buffers.len() < self.max_buffers {
self.free_buffers.push_back(buffer);
}
// Otherwise, buffer is dropped and freed
}
/// Clear all cached buffers
pub(super) fn clear(&mut self) {
self.free_buffers.clear();
}
}
/// GPU Memory Pool - manages reusable GPU memory buffers
pub struct GpuMemoryPool {
device: Arc<CudaContext>,
/// Buckets for i32 buffers (key = size)
i32_buckets: Arc<Mutex<HashMap<usize, MemoryBucket<i32>>>>,
/// Buckets for i64 buffers (key = size)
i64_buckets: Arc<Mutex<HashMap<usize, MemoryBucket<i64>>>>,
/// Buckets for f32 buffers (key = size)
f32_buckets: Arc<Mutex<HashMap<usize, MemoryBucket<f32>>>>,
/// Maximum total memory to cache (bytes)
max_cache_bytes: usize,
/// Current cached memory (bytes)
cached_bytes: Arc<Mutex<usize>>,
}
impl GpuMemoryPool {
/// Create a new memory pool
pub fn new(device: Arc<CudaContext>, max_cache_mb: usize) -> Self {
Self {
device,
i32_buckets: Arc::new(Mutex::new(HashMap::new())),
i64_buckets: Arc::new(Mutex::new(HashMap::new())),
f32_buckets: Arc::new(Mutex::new(HashMap::new())),
max_cache_bytes: max_cache_mb * 1024 * 1024,
cached_bytes: Arc::new(Mutex::new(0)),
}
}
/// Round up size to next bucket size
fn round_to_bucket_size(size: usize) -> usize {
const BUCKET_SIZES: &[usize] = &[
256, // 256 elements
1024, // 1K
4096, // 4K
16384, // 16K
65536, // 64K
262144, // 256K
1048576, // 1M
];
for &bucket_size in BUCKET_SIZES {
if size <= bucket_size {
return bucket_size;
}
}
// For very large sizes, round up to nearest 1M
((size + 1048575) / 1048576) * 1048576
}
/// Allocate i32 buffer (from pool or new)
pub fn alloc_i32(
&self,
size: usize,
) -> Result<CudaSlice<i32>, cudarc::driver::DriverError> {
let bucket_size = Self::round_to_bucket_size(size);
// Try to get from pool
{
let mut buckets = self.i32_buckets.lock().unwrap();
if let Some(bucket) = buckets.get_mut(&bucket_size) {
if let Some(mut buffer) = bucket.try_alloc() {
// Update cached bytes
let mut cached = self.cached_bytes.lock().unwrap();
*cached = cached.saturating_sub(bucket_size * std::mem::size_of::<i32>());
// Resize if needed
if buffer.len() != size {
buffer = self.device.default_stream().alloc_zeros::<i32>(size)?;
}
return Ok(buffer);
}
}
}
// Allocate new buffer
self.device.default_stream().alloc_zeros::<i32>(size)
}
/// Allocate i64 buffer (from pool or new)
pub fn alloc_i64(
&self,
size: usize,
) -> Result<CudaSlice<i64>, cudarc::driver::DriverError> {
let bucket_size = Self::round_to_bucket_size(size);
// Try to get from pool
{
let mut buckets = self.i64_buckets.lock().unwrap();
if let Some(bucket) = buckets.get_mut(&bucket_size) {
if let Some(mut buffer) = bucket.try_alloc() {
// Update cached bytes
let mut cached = self.cached_bytes.lock().unwrap();
*cached = cached.saturating_sub(bucket_size * std::mem::size_of::<i64>());
// Resize if needed
if buffer.len() != size {
buffer = self.device.default_stream().alloc_zeros::<i64>(size)?;
}
return Ok(buffer);
}
}
}
// Allocate new buffer
self.device.default_stream().alloc_zeros::<i64>(size)
}
/// Allocate f32 buffer (from pool or new)
pub fn alloc_f32(
&self,
size: usize,
) -> Result<CudaSlice<f32>, cudarc::driver::DriverError> {
let bucket_size = Self::round_to_bucket_size(size);
// Try to get from pool
{
let mut buckets = self.f32_buckets.lock().unwrap();
if let Some(bucket) = buckets.get_mut(&bucket_size) {
if let Some(mut buffer) = bucket.try_alloc() {
// Update cached bytes
let mut cached = self.cached_bytes.lock().unwrap();
*cached = cached.saturating_sub(bucket_size * std::mem::size_of::<f32>());
// Resize if needed
if buffer.len() != size {
buffer = self.device.default_stream().alloc_zeros::<f32>(size)?;
}
return Ok(buffer);
}
}
}
// Allocate new buffer
self.device.default_stream().alloc_zeros::<f32>(size)
}
/// Return i32 buffer to pool
pub fn free_i32(&self, buffer: CudaSlice<i32>) {
let size = buffer.len();
let bucket_size = Self::round_to_bucket_size(size);
let byte_size = bucket_size * std::mem::size_of::<i32>();
// Check if we have room in cache
{
let cached = self.cached_bytes.lock().unwrap();
if *cached + byte_size > self.max_cache_bytes {
// Cache full, drop buffer
return;
}
}
// Add to pool
let mut buckets = self.i32_buckets.lock().unwrap();
let bucket = buckets.entry(bucket_size).or_insert_with(|| {
MemoryBucket::new(bucket_size, 8) // Max 8 buffers per bucket
});
bucket.free(buffer);
// Update cached bytes
let mut cached = self.cached_bytes.lock().unwrap();
*cached += byte_size;
}
/// Return i64 buffer to pool
pub fn free_i64(&self, buffer: CudaSlice<i64>) {
let size = buffer.len();
let bucket_size = Self::round_to_bucket_size(size);
let byte_size = bucket_size * std::mem::size_of::<i64>();
// Check if we have room in cache
{
let cached = self.cached_bytes.lock().unwrap();
if *cached + byte_size > self.max_cache_bytes {
return;
}
}
// Add to pool
let mut buckets = self.i64_buckets.lock().unwrap();
let bucket = buckets
.entry(bucket_size)
.or_insert_with(|| MemoryBucket::new(bucket_size, 8));
bucket.free(buffer);
// Update cached bytes
let mut cached = self.cached_bytes.lock().unwrap();
*cached += byte_size;
}
/// Return f32 buffer to pool
pub fn free_f32(&self, buffer: CudaSlice<f32>) {
let size = buffer.len();
let bucket_size = Self::round_to_bucket_size(size);
let byte_size = bucket_size * std::mem::size_of::<f32>();
// Check if we have room in cache
{
let cached = self.cached_bytes.lock().unwrap();
if *cached + byte_size > self.max_cache_bytes {
return;
}
}
// Add to pool
let mut buckets = self.f32_buckets.lock().unwrap();
let bucket = buckets
.entry(bucket_size)
.or_insert_with(|| MemoryBucket::new(bucket_size, 8));
bucket.free(buffer);
// Update cached bytes
let mut cached = self.cached_bytes.lock().unwrap();
*cached += byte_size;
}
/// Clear all cached buffers
pub fn clear(&self) {
self.i32_buckets.lock().unwrap().clear();
self.i64_buckets.lock().unwrap().clear();
self.f32_buckets.lock().unwrap().clear();
*self.cached_bytes.lock().unwrap() = 0;
}
/// Get current cached memory size (bytes)
pub fn cached_bytes(&self) -> usize {
*self.cached_bytes.lock().unwrap()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_bucket_size_rounding() {
assert_eq!(GpuMemoryPool::round_to_bucket_size(100), 256);
assert_eq!(GpuMemoryPool::round_to_bucket_size(256), 256);
assert_eq!(GpuMemoryPool::round_to_bucket_size(257), 1024);
assert_eq!(GpuMemoryPool::round_to_bucket_size(65536), 65536);
assert_eq!(GpuMemoryPool::round_to_bucket_size(100000), 262144);
}
}
}
#[cfg(feature = "gpu")]
pub use gpu_pool::GpuMemoryPool;