1 | /******************************************************************************* |
2 | * Copyright 2020-2022 Intel Corporation |
3 | * |
4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
5 | * you may not use this file except in compliance with the License. |
6 | * You may obtain a copy of the License at |
7 | * |
8 | * http://www.apache.org/licenses/LICENSE-2.0 |
9 | * |
10 | * Unless required by applicable law or agreed to in writing, software |
11 | * distributed under the License is distributed on an "AS IS" BASIS, |
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | * See the License for the specific language governing permissions and |
14 | * limitations under the License. |
15 | *******************************************************************************/ |
16 | |
17 | #include "dnnl_test_common.hpp" |
18 | #include "gtest/gtest.h" |
19 | |
20 | #include "dnnl.hpp" |
21 | |
22 | namespace dnnl { |
23 | |
24 | TEST(primitive_cache_mt_test, TestGeneralCase) { |
25 | using tag = memory::format_tag; |
26 | using dt = memory::data_type; |
27 | |
28 | engine eng(get_test_engine_kind(), 0); |
29 | |
30 | // Flush the cache |
31 | set_primitive_cache_capacity(0); |
32 | set_primitive_cache_capacity(1024); |
33 | |
34 | memory::dim n_primitives = 12; |
35 | |
36 | dnnl::impl::parallel_nd(n_primitives, [&](memory::dim np) { |
37 | auto md = memory::desc({{np, 1, 1, 1}, dt::f32, tag::nchw}); |
38 | auto relu_pd = eltwise_forward::primitive_desc(eng, |
39 | prop_kind::forward_inference, algorithm::eltwise_relu, md, md, |
40 | 0.f); |
41 | auto relu = eltwise_forward(relu_pd); |
42 | }); |
43 | |
44 | ASSERT_EQ(get_primitive_cache_size(), n_primitives); |
45 | } |
46 | |
47 | TEST(primitive_cache_mt_test, TestNestedCase) { |
48 | using tag = memory::format_tag; |
49 | using dt = memory::data_type; |
50 | |
51 | engine eng(get_test_engine_kind(), 0); |
52 | |
53 | // Flush the cache |
54 | set_primitive_cache_capacity(0); |
55 | set_primitive_cache_capacity(1024); |
56 | |
57 | memory::dim n_primitives = 12; |
58 | memory::dim n_srcs = 32; |
59 | |
60 | dnnl::impl::parallel_nd(n_primitives, [&](memory::dim np) { |
61 | std::vector<memory::desc> src_mds(n_srcs); |
62 | std::vector<float> scales(n_srcs, 1.0); |
63 | |
64 | for (memory::dim ns = 0; ns < n_srcs; ++ns) { |
65 | src_mds[ns] = memory::desc({{128, 128}, dt::f32, tag::nc}); |
66 | } |
67 | auto sum_pd = sum::primitive_desc(eng, scales, src_mds); |
68 | auto sum_prim = sum(sum_pd); |
69 | }); |
70 | } |
71 | |
72 | TEST(primitive_cache_mt_test, TestMTCacheHit) { |
73 | using tag = memory::format_tag; |
74 | using dt = memory::data_type; |
75 | |
76 | engine eng(get_test_engine_kind(), 0); |
77 | |
78 | // Flush the cache |
79 | dnnl::set_primitive_cache_capacity(0); |
80 | dnnl::set_primitive_cache_capacity(1024); |
81 | |
82 | int n_primitives = 10; |
83 | |
84 | auto create_eltwise_primitive = [&](int np) { |
85 | auto md = memory::desc({{np, 1, 1, 1}, dt::f32, tag::nchw}); |
86 | auto relu_pd = eltwise_forward::primitive_desc(eng, |
87 | prop_kind::forward_inference, algorithm::eltwise_relu, md, md, |
88 | 0.f); |
89 | auto relu = eltwise_forward(relu_pd); |
90 | }; |
91 | |
92 | // Fill the cache with n_primitives (cache_miss) |
93 | for (int i = 0; i < n_primitives; i++) |
94 | create_eltwise_primitive(i); |
95 | |
96 | // This section should only perform cache_hits |
97 | dnnl::impl::parallel(0, [&](int ithr, int nthr) { |
98 | for (int i = 0; i < n_primitives; i++) |
99 | create_eltwise_primitive(i); |
100 | }); |
101 | |
102 | ASSERT_EQ(get_primitive_cache_size(), n_primitives); |
103 | } |
104 | |
105 | } // namespace dnnl |
106 | |