-
Notifications
You must be signed in to change notification settings - Fork 251
Expand file tree
/
Copy pathresample.cpp
More file actions
176 lines (137 loc) · 7.63 KB
/
Copy pathresample.cpp
File metadata and controls
176 lines (137 loc) · 7.63 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
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
/*
* SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: MIT
*/
#include <catch2/catch_test_macros.hpp>
#include "../utils/helpers.h"
#include <cudnn_frontend.h>
TEST_CASE("Resample Max Pooling NHWC Inference", "[resample][pooling][max][graph]") {
namespace fe = cudnn_frontend;
// This example shows running max pooling graphs when in inference mode.
// See details about support surface in
// https://docs.nvidia.com/deeplearning/cudnn/developer/graph-api.html#resamplefwd
constexpr int N = 8;
constexpr int H = 56;
constexpr int W = 56;
constexpr int C = 8;
fe::graph::Graph graph{};
graph.set_io_data_type(fe::DataType_t::HALF).set_compute_data_type(fe::DataType_t::FLOAT);
auto X = graph.tensor(fe::graph::Tensor_attributes().set_dim({N, C, H, W}).set_stride({H * W * C, 1, W * C, C}));
auto [Y, Index] = graph.resample(X,
fe::graph::Resample_attributes()
.set_generate_index(false)
.set_resampling_mode(fe::ResampleMode_t::MAXPOOL)
.set_padding_mode(fe::PaddingMode_t::NEG_INF_PAD)
.set_window({2, 3})
.set_stride({4, 5})
.set_pre_padding({2, 3})
.set_post_padding({4, 5}));
Y->set_output(true);
assert(Index == nullptr);
// Create a unique_ptr for the cuDNN handle
auto handle_ptr = create_cudnn_handle();
auto handle = *handle_ptr;
REQUIRE(graph.validate().is_good());
REQUIRE(graph.build_operation_graph(handle).is_good());
REQUIRE(graph.create_execution_plans({fe::HeurMode_t::A}).is_good());
REQUIRE(graph.check_support().is_good());
REQUIRE(graph.build_plans(fe::BuildPlanPolicy_t::HEURISTICS_CHOICE).is_good());
Surface<half> X_gpu(N * H * W * C);
Surface<half> Y_gpu(N * H * W * C);
std::unordered_map<std::shared_ptr<fe::graph::Tensor_attributes>, void*> variant_pack = {{X, X_gpu.devPtr},
{Y, Y_gpu.devPtr}};
int64_t workspace_size = 0;
REQUIRE(graph.get_workspace_size(workspace_size).is_good());
Surface<int8_t> workspace(workspace_size);
REQUIRE(graph.execute(handle, variant_pack, workspace.devPtr).is_good());
}
TEST_CASE("Resample Max Pooling NHWC Training", "[resample][pooling][max][graph]") {
namespace fe = cudnn_frontend;
// This example shows running NHWC max pooling graphs.
// Support for NHWC max pooling has a fast path which can dump index tensor from forward pass.
// This mean backward pass to skip reading full X tensor and instead just use this index tensor.
// See details about support surface and index tensor in
// https://docs.nvidia.com/deeplearning/cudnn/developer/graph-api.html#resamplefwd
constexpr int N = 8;
constexpr int H = 56;
constexpr int W = 56;
constexpr int C = 8;
fe::graph::Graph graph{};
graph.set_io_data_type(fe::DataType_t::HALF).set_compute_data_type(fe::DataType_t::FLOAT);
auto X = graph.tensor(fe::graph::Tensor_attributes().set_dim({N, C, H, W}).set_stride({H * W * C, 1, W * C, C}));
auto [Y, Index] = graph.resample(X,
fe::graph::Resample_attributes()
.set_generate_index(true)
.set_resampling_mode(fe::ResampleMode_t::MAXPOOL)
.set_padding_mode(fe::PaddingMode_t::NEG_INF_PAD)
.set_window({2, 3})
.set_stride({4, 5})
.set_pre_padding({2, 3})
.set_post_padding({4, 5}));
Y->set_output(true);
Index->set_output(true).set_data_type(fe::DataType_t::INT8);
// Create a unique_ptr for the cuDNN handle
auto handle_ptr = create_cudnn_handle();
auto handle = *handle_ptr;
REQUIRE(graph.validate().is_good());
auto const status = graph.build_operation_graph(handle);
if (cudnn_frontend::detail::get_backend_version() >= 8600)
REQUIRE(status.is_good());
else {
REQUIRE(status.is_bad());
SKIP("Using index tensor is not supported pre 8.6.");
}
REQUIRE(graph.create_execution_plans({fe::HeurMode_t::A}).is_good());
REQUIRE(graph.check_support().is_good());
REQUIRE(graph.build_plans(fe::BuildPlanPolicy_t::HEURISTICS_CHOICE).is_good());
Surface<half> X_gpu(N * H * W * C);
Surface<half> Y_gpu(N * H * W * C);
Surface<int8_t> Index_gpu(N * H * W * C / 8);
std::unordered_map<std::shared_ptr<fe::graph::Tensor_attributes>, void*> variant_pack = {
{X, X_gpu.devPtr}, {Y, Y_gpu.devPtr}, {Index, Index_gpu.devPtr}};
int64_t workspace_size = 0;
REQUIRE(graph.get_workspace_size(workspace_size).is_good());
Surface<int8_t> workspace(workspace_size);
REQUIRE(graph.execute(handle, variant_pack, workspace.devPtr).is_good());
}
TEST_CASE("Resample Avg Pooling", "[resample][pooling][average][graph]") {
namespace fe = cudnn_frontend;
// This example shows running average pooling graphs.
// There is no difference between NHWC and NCHW support surface.
// See details about support surface in
// https://docs.nvidia.com/deeplearning/cudnn/developer/graph-api.html#resamplefwd
constexpr int N = 8;
constexpr int H = 56;
constexpr int W = 56;
constexpr int C = 8;
fe::graph::Graph graph{};
graph.set_io_data_type(fe::DataType_t::HALF).set_compute_data_type(fe::DataType_t::FLOAT);
auto X = graph.tensor(fe::graph::Tensor_attributes().set_dim({N, C, H, W}).set_stride({H * W * C, 1, W * C, C}));
auto [Y, Index] = graph.resample(X,
fe::graph::Resample_attributes()
.set_generate_index(true)
.set_resampling_mode(fe::ResampleMode_t::AVGPOOL_INCLUDE_PADDING)
.set_padding_mode(fe::PaddingMode_t::ZERO_PAD)
.set_window({2, 3})
.set_stride({4, 5})
.set_pre_padding({2, 3})
.set_post_padding({4, 5}));
Y->set_output(true);
assert(Index == nullptr);
// Create a unique_ptr for the cuDNN handle
auto handle_ptr = create_cudnn_handle();
auto handle = *handle_ptr;
REQUIRE(graph.validate().is_good());
REQUIRE(graph.build_operation_graph(handle).is_good());
REQUIRE(graph.create_execution_plans({fe::HeurMode_t::A}).is_good());
REQUIRE(graph.check_support().is_good());
REQUIRE(graph.build_plans(fe::BuildPlanPolicy_t::HEURISTICS_CHOICE).is_good());
Surface<half> X_gpu(N * H * W * C);
Surface<half> Y_gpu(N * H * W * C);
std::unordered_map<std::shared_ptr<fe::graph::Tensor_attributes>, void*> variant_pack = {{X, X_gpu.devPtr},
{Y, Y_gpu.devPtr}};
int64_t workspace_size = 0;
REQUIRE(graph.get_workspace_size(workspace_size).is_good());
Surface<int8_t> workspace(workspace_size);
REQUIRE(graph.execute(handle, variant_pack, workspace.devPtr).is_good());
}