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
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// SPDX-FileCopyrightText: Copyright Contributors to the Kokkos project
#include <TestCuda_Category.hpp>
#include <TestMultiGPU.hpp>
namespace {
struct StreamsAndDevices {
std::array<cudaStream_t, 2> streams;
std::array<int, 2> devices;
StreamsAndDevices() {
int n_devices;
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaGetDeviceCount(&n_devices));
devices = {0, n_devices - 1};
for (int i = 0; i < 2; ++i) {
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaSetDevice(devices[i]));
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaStreamCreate(&streams[i]));
}
}
StreamsAndDevices(const StreamsAndDevices &) = delete;
StreamsAndDevices &operator=(const StreamsAndDevices &) = delete;
~StreamsAndDevices() {
for (int i = 0; i < 2; ++i) {
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaSetDevice(devices[i]));
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaStreamDestroy(streams[i]));
}
}
};
std::array<TEST_EXECSPACE, 2> get_execution_spaces(
const StreamsAndDevices &streams_and_devices) {
TEST_EXECSPACE exec0(streams_and_devices.streams[0]);
TEST_EXECSPACE exec1(streams_and_devices.streams[1]);
// Must return void to use ASSERT_EQ
[&]() {
ASSERT_EQ(exec0.cuda_device(), streams_and_devices.devices[0]);
ASSERT_EQ(exec1.cuda_device(), streams_and_devices.devices[1]);
}();
return {exec0, exec1};
}
struct TEST_CATEGORY_FIXTURE(multi_gpu) : public ::testing::Test {
StreamsAndDevices sd;
void SetUp() override {
if (sd.devices[0] == sd.devices[1])
GTEST_SKIP() << "Skipping Cuda multi-gpu testing since current machine "
"only contains a single GPU.\n";
}
};
TEST_F(TEST_CATEGORY_FIXTURE(multi_gpu), managed_views) {
StreamsAndDevices streams_and_devices;
{
std::array<TEST_EXECSPACE, 2> execs =
get_execution_spaces(streams_and_devices);
Kokkos::View<int *, TEST_EXECSPACE> view0(
Kokkos::view_alloc("v0", execs[0]), 100);
Kokkos::View<int *, TEST_EXECSPACE> view(Kokkos::view_alloc("v", execs[1]),
100);
test_policies(execs[0], view0, execs[1], view);
}
}
TEST_F(TEST_CATEGORY_FIXTURE(multi_gpu), unmanaged_views) {
StreamsAndDevices streams_and_devices;
{
std::array<TEST_EXECSPACE, 2> execs =
get_execution_spaces(streams_and_devices);
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaSetDevice(execs[0].cuda_device()));
int *p0;
void *p0_void_ptr = nullptr;
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaMalloc(&p0_void_ptr, sizeof(int) * 100));
p0 = static_cast<int *>(p0_void_ptr);
Kokkos::View<int *, TEST_EXECSPACE> view0(p0, 100);
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaSetDevice(execs[1].cuda_device()));
int *p;
void *p_void_ptr = nullptr;
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaMalloc(&p_void_ptr, sizeof(int) * 100));
p = static_cast<int *>(p_void_ptr);
Kokkos::View<int *, TEST_EXECSPACE> view(p, 100);
test_policies(execs[0], view0, execs[1], view);
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaFree(p0));
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaFree(p));
}
}
TEST_F(TEST_CATEGORY_FIXTURE(multi_gpu), scratch_space) {
StreamsAndDevices streams_and_devices;
{
std::array<TEST_EXECSPACE, 2> execs =
get_execution_spaces(streams_and_devices);
test_scratch(execs[0], execs[1]);
}
}
TEST_F(TEST_CATEGORY_FIXTURE(multi_gpu), stream_sync_semantics_raw_cuda) {
// Test that stream synchronization behavior for various GPU APIs matches the
// assumptions made in Kokkos for multi gpu support, namely, that any stream
// (no matter which device it is created on) can be synced from any device.
StreamsAndDevices streams_and_devices;
{
auto streams = streams_and_devices.streams;
auto devices = streams_and_devices.devices;
// Allocate data.
int *value;
void *value_void_ptr = nullptr;
KOKKOS_IMPL_CUDA_SAFE_CALL(
cudaMallocHost(&value_void_ptr, 1 * sizeof(int)));
value = static_cast<int *>(value_void_ptr);
int *check;
void *check_void_ptr = nullptr;
KOKKOS_IMPL_CUDA_SAFE_CALL(
cudaMallocHost(&check_void_ptr, 1 * sizeof(int)));
check = static_cast<int *>(check_void_ptr);
// Launch "long" kernel on device 0.
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaSetDevice(devices[0]));
constexpr size_t size = 10000;
accumulate_kernel<size><<<1, 1, 0, streams[0]>>>(value);
// Wait for the kernel running on device 0 while we are on device 1, then
// check the value.
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaSetDevice(devices[1]));
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaStreamSynchronize(streams[0]));
copy_kernel<<<1, 1, 0, streams[1]>>>(check, value);
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaStreamSynchronize(streams[1]));
ASSERT_EQ(check[0], size);
// Cleanup.
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaFreeHost(value));
KOKKOS_IMPL_CUDA_SAFE_CALL(cudaFreeHost(check));
}
}
} // namespace
|
