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
// crates/xlog-cuda/tests/test_runtime_direct_allocator_parallel_stress.rs
//! Regression coverage for the v0.6 device-runtime allocator's
//! direct-backend parallel allocation stress contract.
//!
//! Per the locked acceptance criteria: 8–16 host threads each
//! repeatedly allocate device blocks of varied sizes via
//! `XlogDeviceRuntime::try_get(0).allocate(...)`, write a
//! deterministic byte pattern on the GPU, read it back, verify it
//! matches, then deallocate. After all threads finish,
//! `bytes_outstanding()` must return to its baseline.
//!
//! Status: this test is expected to **pass** on `DirectCudaResource`
//! (the cudarc direct allocation backend). cudarc's synchronous
//! `CudaDevice::alloc::<u8>` and the underlying `cuMemAlloc`/
//! `cuMemFree` are device-wide and not stream-ordered, so parallel
//! callers should not see overlapping allocations or corrupted
//! bookkeeping. A failure indicates a real bug in the direct backend
//! or the singleton — stop and debug before proceeding to the
//! stream-ordered async-backend checks.
//!
//! This is a sanity check for the singleton + resource layer, not a
//! demonstration of stream-ordered correctness. The async-backend
//! stream-ordering contract is covered by the dedicated
//! stream-ordered allocation lifetime tests.
//!
//! The byte-pattern check goes through `cuMemcpyHtoD_v2` /
//! `cuMemcpyDtoH_v2` on the raw `DeviceBlock::ptr`. This is the
//! lowest-level API that exercises the block's memory directly,
//! independent of any cudarc safe wrapper, so a cross-thread pointer
//! aliasing bug would surface as a content mismatch rather than a
//! benign bookkeeping success.
use std::sync::{Arc, Barrier};
use std::thread;
use cudarc::driver::sys;
use xlog_cuda::device_runtime::{AllocTag, StreamId, XlogDeviceRuntime};
const NUM_THREADS: usize = 8;
const ITERATIONS_PER_THREAD: usize = 64;
/// Tiny LCG so we don't pull in `rand`. Per-thread seed is
/// `(thread_idx + 1) * 0x9E3779B97F4A7C15`. Deterministic across runs
/// for a fixed `NUM_THREADS`.
#[derive(Clone, Copy)]
struct Lcg(u64);
impl Lcg {
fn new(seed: u64) -> Self {
Self(seed.max(1))
}
fn next_u64(&mut self) -> u64 {
self.0 = self
.0
.wrapping_mul(6364136223846793005)
.wrapping_add(1442695040888963407);
self.0
}
fn next_in(&mut self, lo: usize, hi: usize) -> usize {
let span = (hi - lo) as u64;
lo + (self.next_u64() % span) as usize
}
}
unsafe fn htod(ptr: u64, host: &[u8]) {
let res = sys::cuMemcpyHtoD_v2(ptr, host.as_ptr() as *const _, host.len());
assert_eq!(
res,
sys::cudaError_enum::CUDA_SUCCESS,
"cuMemcpyHtoD_v2 returned {:?} for {} bytes at {:#x}",
res,
host.len(),
ptr
);
}
unsafe fn dtoh(ptr: u64, host: &mut [u8]) {
let res = sys::cuMemcpyDtoH_v2(host.as_mut_ptr() as *mut _, ptr, host.len());
assert_eq!(
res,
sys::cudaError_enum::CUDA_SUCCESS,
"cuMemcpyDtoH_v2 returned {:?} for {} bytes at {:#x}",
res,
host.len(),
ptr
);
}
#[test]
fn direct_allocator_parallel_stress_alloc_write_verify_dealloc() {
let runtime = match XlogDeviceRuntime::try_get(0) {
Ok(rt) => rt,
Err(err) => {
eprintln!(
"Skipping direct allocator parallel stress: CUDA runtime unavailable: {}",
err
);
return;
}
};
let baseline = runtime.bytes_outstanding();
let barrier = Arc::new(Barrier::new(NUM_THREADS));
let mut handles = Vec::with_capacity(NUM_THREADS);
for thread_idx in 0..NUM_THREADS {
let barrier = Arc::clone(&barrier);
let runtime_ref = runtime;
handles.push(thread::spawn(move || {
// Wait until all threads are ready so the contention
// window starts together.
barrier.wait();
let seed: u64 = (thread_idx as u64 + 1).wrapping_mul(0x9E37_79B9_7F4A_7C15);
let mut rng = Lcg::new(seed);
// Per-thread tag pool: tags must be 'static; pre-pick
// from a small const set indexed by thread.
let tag = match thread_idx % 4 {
0 => AllocTag("direct-thread-0"),
1 => AllocTag("direct-thread-1"),
2 => AllocTag("direct-thread-2"),
_ => AllocTag("direct-thread-other"),
};
for iter in 0..ITERATIONS_PER_THREAD {
let bytes = rng.next_in(64, 4096 + 1);
let block = runtime_ref
.allocate(bytes, StreamId::DEFAULT, tag)
.unwrap_or_else(|e| {
panic!("thread {} iter {}: allocate({}) failed: {}", thread_idx, iter, bytes, e)
});
assert_eq!(block.bytes, bytes);
assert_eq!(block.device_ordinal, 0);
// Deterministic per-(thread, iter) byte pattern: a
// pointer-aliasing bug would write thread A's bytes
// into thread B's block and the readback assertion
// below would fail.
let stamp: u8 = ((thread_idx as u32).wrapping_add(iter as u32) & 0xFF) as u8;
let host_in: Vec<u8> = (0..bytes)
.map(|i| stamp.wrapping_add((i & 0xFF) as u8))
.collect();
let mut host_out = vec![0u8; bytes];
// SAFETY: block.ptr is a live device pointer of size
// `bytes` returned by the runtime's allocator. The
// host_in/host_out slices have matching length. The
// synchronous memcpy variants block until the copy
// completes on the default stream.
unsafe {
htod(block.ptr, &host_in);
dtoh(block.ptr, &mut host_out);
}
assert_eq!(
host_out, host_in,
"thread {} iter {}: byte pattern mismatch — possible cross-thread pointer aliasing",
thread_idx, iter
);
runtime_ref
.deallocate(block)
.unwrap_or_else(|e| panic!("thread {} iter {}: deallocate failed: {}", thread_idx, iter, e));
}
}));
}
for (i, h) in handles.into_iter().enumerate() {
h.join()
.unwrap_or_else(|_| panic!("direct allocator stress thread {} panicked", i));
}
let after = runtime.bytes_outstanding();
assert_eq!(
after, baseline,
"bytes_outstanding leaked: baseline {}, after {} — bookkeeping race in DirectCudaResource",
baseline, after
);
}