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
|
use super::{GpuError, Message, ResultMessage};
use ocl::{flags, Buffer, Context, Device, Kernel, Platform, Program, Queue};
use std::sync::mpsc::{channel, Receiver, Sender};
#[derive(Debug)]
pub struct Host {
#[allow(unused)]
platform: Platform,
#[allow(unused)]
device: Device,
#[allow(unused)]
context: Context,
program: Program,
queue: Queue,
n: u32,
h: u32,
w: u32,
/// Workgroup size, set to 0 for max
wg_size: usize,
permutations: Buffer<u64>,
receiver: Receiver<Message>,
output_sender: Sender<Message>,
}
impl Host {
pub fn launch_sevice(
permutation_masks: &[u64],
n: u32,
h: u32,
w: u32,
wg_size: usize,
src: &str,
output_sender: Sender<Message>,
) -> ocl::Result<(Sender<Message>, std::thread::JoinHandle<()>)> {
let platform = ocl::Platform::default();
let device = ocl::Device::first(platform)?;
let context = ocl::Context::builder()
.platform(platform)
.devices(device.clone())
.build()?;
let queue = ocl::Queue::new(&context, device, None)?;
let program = Program::builder()
.devices(device)
.src(src)
.build(&context)?;
let buffer = ocl::Buffer::builder()
.queue(queue.clone())
.flags(flags::MEM_READ_WRITE)
.copy_host_slice(permutation_masks)
.len(permutation_masks.len())
.build()?;
let (sender, receiver) = channel();
let solver = Self {
platform,
device,
context,
program,
queue,
n,
h,
w,
wg_size,
permutations: buffer,
receiver,
output_sender,
};
let handle = std::thread::Builder::new()
.name("GPU Host Deamon".into())
.spawn(move || {
if let Err(err) = solver.run() {
println!("{}", err);
}
})
.unwrap();
println!("started gpu thread");
Ok((sender, handle))
}
fn get_dim(&self, queue: usize) -> usize {
let chunk = self.permutations.len() / self.n as usize;
let dim = (queue + 1) * chunk;
(dim + self.wg_size - 1) / self.wg_size * self.wg_size
}
fn get_off(&self, queue: usize) -> usize {
let chunk = self.permutations.len() / self.n as usize;
if self.permutations.len() < chunk + self.get_dim(queue) {
panic!("workgroup size too big; offset underflow")
}
self.permutations.len() - chunk - self.get_dim(queue)
}
fn get_res(&self, queue: usize) -> usize {
let dim = self.get_dim(queue);
dim * self.get_res_save_dim()
}
fn get_res_save_dim(&self) -> usize {
(self.wg_size + 63) / 64
}
fn run(self) -> Result<(), GpuError> {
let queues = (self.n - self.h + 1) as usize;
let mut instruction_buffer = Vec::with_capacity((self.n - self.h) as usize);
let mut result_buffer = Vec::with_capacity((self.n - self.h) as usize);
for i in 0..queues {
let buffer: Buffer<u64> = Buffer::builder()
.queue(self.queue.clone())
.len(self.wg_size)
.flags(flags::MEM_READ_WRITE)
.build()?;
instruction_buffer.push(buffer);
let results: Buffer<u64> = Buffer::builder()
.queue(self.queue.clone())
.len(self.get_res(i))
.flags(flags::MEM_READ_WRITE)
.build()?;
result_buffer.push(results);
}
println!("finished gpu setup");
loop {
match self.receiver.recv()? {
Message::CpuDone => {
self.output_sender.send(Message::CpuDone)?;
return Ok(());
}
Message::Terminate => {
return Ok(());
}
Message::HostMessage((id, i, buffer)) => {
let i = i as usize;
let off = self.get_off(i);
let dim = self.get_dim(i);
let res = self.get_res(i);
let res_size = self.get_res_save_dim();
instruction_buffer[i].write(&buffer).enq()?;
//println!("dim: {}", dim);
//println!("off: {}", self.get_off(i));
//println!("result size: {}", self.get_res_save_dim());
let kernel = Kernel::builder()
.program(&self.program)
.name("check")
.queue(self.queue.clone())
.global_work_size(dim)
.arg(&self.permutations)
.arg(&result_buffer[i])
.arg(&instruction_buffer[i])
.arg_local::<u64>(self.wg_size)
.arg(self.n)
.arg(self.w)
.arg(off)
.build()?;
unsafe {
kernel
.cmd()
.queue(&self.queue)
.global_work_offset(kernel.default_global_work_offset())
.global_work_size(dim)
.local_work_size(self.wg_size)
.enq()?;
}
// (5) Read results from the device into a vector (`::block` not shown):
let mut data = vec![0u64; res];
result_buffer[i]
.cmd()
.queue(&self.queue)
.offset(0)
.read(&mut data)
.enq()?;
self.queue.finish()?;
self.output_sender
.send(Message::ResultMessage(ResultMessage::new(
data,
off,
res_size,
self.wg_size,
id,
)))?;
}
m => println!("Invalid MessageType {:?} recived by host", m),
}
}
}
}
|
