cuda.qbk
4.21 KB
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
[/
Copyright Oliver Kowalke 2017.
Distributed under the Boost Software License, Version 1.0.
(See accompanying file LICENSE_1_0.txt or copy at
http://www.boost.org/LICENSE_1_0.txt
]
[#cuda]
[section:cuda CUDA]
[@http://developer.nvidia.com/cuda-zone/ CUDA (Compute Unified Device Architecture)] is a platform for parallel computing
on NVIDIA GPUs. The application programming interface of CUDA gives access to
GPU's instruction set and computation resources (Execution of compute kernels).
[heading Synchronization with CUDA streams]
CUDA operation such as compute kernels or memory transfer (between host and
device) can be grouped/queued by CUDA streams. are executed on the GPUs.
Boost.Fiber enables a fiber to sleep (suspend) till a CUDA stream has completed
its operations. This enables applications to run other fibers on the CPU without
the need to spawn an additional OS-threads. And resume the fiber when the CUDA
streams has finished.
__global__
void kernel( int size, int * a, int * b, int * c) {
int idx = threadIdx.x + blockIdx.x * blockDim.x;
if ( idx < size) {
int idx1 = (idx + 1) % 256;
int idx2 = (idx + 2) % 256;
float as = (a[idx] + a[idx1] + a[idx2]) / 3.0f;
float bs = (b[idx] + b[idx1] + b[idx2]) / 3.0f;
c[idx] = (as + bs) / 2;
}
}
boost::fibers::fiber f([&done]{
cudaStream_t stream;
cudaStreamCreate( & stream);
int size = 1024 * 1024;
int full_size = 20 * size;
int * host_a, * host_b, * host_c;
cudaHostAlloc( & host_a, full_size * sizeof( int), cudaHostAllocDefault);
cudaHostAlloc( & host_b, full_size * sizeof( int), cudaHostAllocDefault);
cudaHostAlloc( & host_c, full_size * sizeof( int), cudaHostAllocDefault);
int * dev_a, * dev_b, * dev_c;
cudaMalloc( & dev_a, size * sizeof( int) );
cudaMalloc( & dev_b, size * sizeof( int) );
cudaMalloc( & dev_c, size * sizeof( int) );
std::minstd_rand generator;
std::uniform_int_distribution<> distribution(1, 6);
for ( int i = 0; i < full_size; ++i) {
host_a[i] = distribution( generator);
host_b[i] = distribution( generator);
}
for ( int i = 0; i < full_size; i += size) {
cudaMemcpyAsync( dev_a, host_a + i, size * sizeof( int), cudaMemcpyHostToDevice, stream);
cudaMemcpyAsync( dev_b, host_b + i, size * sizeof( int), cudaMemcpyHostToDevice, stream);
kernel<<< size / 256, 256, 0, stream >>>( size, dev_a, dev_b, dev_c);
cudaMemcpyAsync( host_c + i, dev_c, size * sizeof( int), cudaMemcpyDeviceToHost, stream);
}
auto result = boost::fibers::cuda::waitfor_all( stream); // suspend fiber till CUDA stream has finished
BOOST_ASSERT( stream == std::get< 0 >( result) );
BOOST_ASSERT( cudaSuccess == std::get< 1 >( result) );
std::cout << "f1: GPU computation finished" << std::endl;
cudaFreeHost( host_a);
cudaFreeHost( host_b);
cudaFreeHost( host_c);
cudaFree( dev_a);
cudaFree( dev_b);
cudaFree( dev_c);
cudaStreamDestroy( stream);
});
f.join();
[heading Synopsis]
#include <boost/fiber/cuda/waitfor.hpp>
namespace boost {
namespace fibers {
namespace cuda {
std::tuple< cudaStream_t, cudaError_t > waitfor_all( cudaStream_t st);
std::vector< std::tuple< cudaStream_t, cudaError_t > > waitfor_all( cudaStream_t ... st);
}}}
[ns_function_heading cuda..waitfor]
#include <boost/fiber/cuda/waitfor.hpp>
namespace boost {
namespace fibers {
namespace cuda {
std::tuple< cudaStream_t, cudaError_t > waitfor_all( cudaStream_t st);
std::vector< std::tuple< cudaStream_t, cudaError_t > > waitfor_all( cudaStream_t ... st);
}}}
[variablelist
[[Effects:] [Suspends active fiber till CUDA stream has finished its operations.]]
[[Returns:] [tuple of stream reference and the CUDA stream status]]
]
[endsect]