1 // Copyright 2022 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <algorithm> // For std::generate.
7 #include <array> // For std::array.
8 #include <cstddef> // For size_t.
9 #include <functional> // For std::multiplies.
10 #include <memory> // For std::unique_ptr.
11 #include <random> // For std::random_device, std::mt19937, std::uniform_real_distribution.
12 #include <vector> // For std::vector.
13
14 #include <xnnpack.h>
15 #include <xnnpack/node-type.h>
16 #include <xnnpack/operator.h>
17 #include <xnnpack/subgraph.h>
18
19 #include "subgraph-unary-tester.h"
20 #include <gtest/gtest.h>
21
22 using HardSwishTestF32 = UnaryTest<float>;
23
TEST_F(HardSwishTestF32,define)24 TEST_F(HardSwishTestF32, define)
25 {
26 ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
27
28 xnn_subgraph_t subgraph = nullptr;
29 ASSERT_EQ(xnn_status_success, xnn_create_subgraph(0, /*flags=*/0, &subgraph));
30 std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);
31 const std::array<size_t, 3> dims = {{1, 3, 5}};
32 uint32_t input_id = XNN_INVALID_NODE_ID;
33 ASSERT_EQ(
34 xnn_status_success,
35 xnn_define_tensor_value(
36 subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, XNN_INVALID_VALUE_ID, /*flags=*/0, &input_id));
37 ASSERT_NE(input_id, XNN_INVALID_NODE_ID);
38
39 uint32_t output_id = XNN_INVALID_NODE_ID;
40 ASSERT_EQ(
41 xnn_status_success,
42 xnn_define_tensor_value(
43 subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, XNN_INVALID_VALUE_ID, /*flags=*/0, &output_id));
44 ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
45
46 ASSERT_EQ(xnn_status_success, xnn_define_hardswish(subgraph, input_id, output_id, 0));
47
48 ASSERT_EQ(subgraph->num_nodes, 1);
49 const struct xnn_node* node = &subgraph->nodes[0];
50 ASSERT_EQ(node->type, xnn_node_type_hardswish);
51 ASSERT_EQ(node->compute_type, xnn_compute_type_fp32);
52 ASSERT_EQ(node->num_inputs, 1);
53 ASSERT_EQ(node->inputs[0], input_id);
54 ASSERT_EQ(node->num_outputs, 1);
55 ASSERT_EQ(node->outputs[0], output_id);
56 ASSERT_EQ(node->flags, 0);
57 }
58
TEST_F(HardSwishTestF32,matches_operator_api)59 TEST_F(HardSwishTestF32, matches_operator_api)
60 {
61 std::vector<float> input(num_output_elements + XNN_EXTRA_BYTES / sizeof(float), std::nanf(""));
62 std::uniform_real_distribution<float> f32dist(-4.0f, 4.0f);
63 std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
64 std::vector<float> subgraph_output(num_output_elements, std::nanf(""));
65
66 ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
67
68 // Call operator API.
69 xnn_operator_t op = nullptr;
70 xnn_status status = xnn_create_hardswish_nc_f32(channels, channels, channels, /*flags=*/0, &op);
71 if (status == xnn_status_unsupported_hardware) {
72 GTEST_SKIP();
73 }
74 ASSERT_EQ(xnn_status_success, status);
75 ASSERT_NE(nullptr, op);
76 std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);
77 std::vector<float> operator_output(num_output_elements, std::nanf(""));
78 ASSERT_EQ(
79 xnn_status_success,
80 xnn_setup_hardswish_nc_f32(op, batch_size, input.data(), operator_output.data(), /*threadpool=*/nullptr));
81 ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));
82
83 // Call subgraph API.
84 xnn_subgraph_t subgraph = nullptr;
85 ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
86 ASSERT_NE(nullptr, subgraph);
87 std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);
88 uint32_t input_id = XNN_INVALID_NODE_ID;
89 ASSERT_EQ(
90 xnn_status_success, xnn_define_tensor_value(
91 subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, /*external_id=*/0,
92 XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
93 ASSERT_NE(input_id, XNN_INVALID_NODE_ID);
94 uint32_t output_id = XNN_INVALID_NODE_ID;
95 ASSERT_EQ(
96 xnn_status_success, xnn_define_tensor_value(
97 subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, /*external_id=*/1,
98 XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
99 ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
100 ASSERT_EQ(xnn_status_success, xnn_define_hardswish(subgraph, input_id, output_id, /*flags=*/0));
101 xnn_runtime_t runtime = nullptr;
102 ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
103 ASSERT_NE(nullptr, runtime);
104 std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
105 std::array<xnn_external_value, 2> external = {
106 xnn_external_value{input_id, input.data()},
107 xnn_external_value{output_id, subgraph_output.data()}
108 };
109 ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
110 ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));
111
112 // Check outputs match.
113 for (size_t i = 0; i < num_output_elements; i++) {
114 ASSERT_EQ(subgraph_output[i], operator_output[i]);
115 }
116 }
117