aboutsummaryrefslogtreecommitdiff
path: root/deps/cudnn-frontend/include/cudnn_frontend/node/reduction.h
blob: 5944b69a33d3b725f32887c18c3dd9c7311af5f0 (plain) (blame)
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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
#pragma once

#include "../graph_helpers.h"
#include "../node_interface.h"

namespace cudnn_frontend::graph {

class ReductionNode : public NodeCRTP<ReductionNode> {
   public:
    Reduction_attributes attributes;

    ReductionNode(Reduction_attributes&& attributes_, detail::Context const& context)
        : NodeCRTP(context), attributes(std::move(attributes_)) {}

    Type
    getType() override final {
        return Type::REDUCTION;
    }

    error_t
    pre_validate_node() const override final {
        CUDNN_FE_LOG_LABEL_ENDL("INFO: Validating ReductionNode " << attributes.name);

        if (attributes.get_is_deterministic() && detail::get_backend_version() < 91100) {
            return {error_code_t::GRAPH_NOT_SUPPORTED, "DETERMINISTIC mode is not supported in cudnn version < 9.11.0"};
        }
        return {error_code_t::OK, ""};
    }

    error_t
    infer_properties_node() override final {
        CUDNN_FE_LOG_LABEL_ENDL("INFO:     Inferrencing properties for reduction node " << attributes.name);

        attributes.fill_from_context(context);

        // Only inferrencing from IN_0 to OUT_0 works today.
        auto x_tensor = attributes.inputs[Reduction_attributes::input_names::X];
        auto y_tensor = attributes.outputs[Reduction_attributes::output_names::Y];

        auto const& x_tensor_dim = x_tensor->get_dim();
        auto y_tensor_dim        = y_tensor->get_dim();
        // Only infer dims and strides if user did not set them
        if (y_tensor_dim.empty()) {
            y_tensor->set_dim(x_tensor_dim);
        }
        if (y_tensor->get_stride().empty()) {
            auto const& y_dim = y_tensor->get_dim();
            // Default to NHWC
            auto const& stride_order = detail::generate_NHWC_stride_order(y_dim.size());
            y_tensor->set_stride(detail::generate_stride(y_dim, stride_order));
        }

        return {error_code_t::OK, ""};
    }

    error_t
    create_cudnn_operations(
        std::unordered_set<Tensor_attributes::uid_t>& uids_involved_in_operations,
        std::vector<std::shared_ptr<cudnn_frontend::Operation>>& operations,
        managed_backend_descriptor_t& raw_operations,
        std::unordered_map<int64_t, std::shared_ptr<cudnn_frontend::Tensor>>& tensors) const override final {
        CUDNN_FRONTEND_UNUSED(raw_operations);
        CUDNN_FE_LOG_LABEL("INFO: " << "Building ReductionNode operations " << attributes.name << " ");

        // Create reduction descriptor by directly calling cuDNN backend API
        ReductionDesc_v8 reduction_descriptor;

        // 1. Create the backend descriptor

        _CUDNN_CHECK_CUDNN_ERROR(
            reduction_descriptor.initialize_managed_backend_pointer(CUDNN_BACKEND_REDUCTION_DESCRIPTOR));

        // 2. Set compute type attribute
        cudnnDataType_t cudnn_data_type;

        _CUDNN_CHECK_CUDNN_ERROR(detail::convert_to_cudnn_type(attributes.compute_data_type, cudnn_data_type));

        _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(reduction_descriptor.get_raw_desc(),
                                                       CUDNN_ATTR_REDUCTION_COMP_TYPE,
                                                       CUDNN_TYPE_DATA_TYPE,
                                                       1,
                                                       &cudnn_data_type));

        // 3. Set reduction operator attribute
        cudnnReduceTensorOp_t cudnn_reduction_mode;

        _CUDNN_CHECK_CUDNN_ERROR(detail::convert_to_cudnn_type(attributes.get_mode().value(), cudnn_reduction_mode));

        _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(reduction_descriptor.get_raw_desc(),
                                                       CUDNN_ATTR_REDUCTION_OPERATOR,
                                                       CUDNN_TYPE_REDUCTION_OPERATOR_TYPE,
                                                       1,
                                                       &cudnn_reduction_mode));

        // 4. Set deterministic mode if supported
#if (CUDNN_VERSION >= 91100)
        if (detail::get_backend_version() >= 91100) {
            bool is_deterministic = attributes.get_is_deterministic();

            _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(reduction_descriptor.get_raw_desc(),
                                                           CUDNN_ATTR_REDUCTION_IS_DETERMINISTIC,
                                                           CUDNN_TYPE_BOOLEAN,
                                                           1,
                                                           &is_deterministic));
        }
#endif

        // 5. Finalize the descriptor

        _CUDNN_CHECK_CUDNN_ERROR(detail::finalize(reduction_descriptor.get_raw_desc()));
        CUDNN_FE_LOG_LABEL_ENDL(reduction_descriptor);

        // Create operation by directly calling cuDNN backend API
        Operation_v8 reduction_operation;

        // Validate input tensors are set
        CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(X, Reduction_attributes::input_names::X);
        auto Group_offset = attributes.inputs.find(Reduction_attributes::input_names::Group_offset);
        CUDNN_FE_VALIDATE_AND_ASSIGN_OUTPUT_TENSOR(Y, Reduction_attributes::output_names::Y);

        // 1. Create the backend operation descriptor

        _CUDNN_CHECK_CUDNN_ERROR(
            reduction_operation.initialize_managed_backend_pointer(CUDNN_BACKEND_OPERATION_REDUCTION_DESCRIPTOR));

        // 2. Set the reduction descriptor attribute
        auto reduction_desc_ptr = reduction_descriptor.get_raw_desc();

        _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(reduction_operation.get_raw_desc(),
                                                       CUDNN_ATTR_OPERATION_REDUCTION_DESC,
                                                       CUDNN_TYPE_BACKEND_DESCRIPTOR,
                                                       1,
                                                       &reduction_desc_ptr));

        // 3. Set the input tensor (X) descriptor attribute
        auto x_backend_desc = tensors.at(X->second->get_uid())->get_raw_desc();

        _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(reduction_operation.get_raw_desc(),
                                                       CUDNN_ATTR_OPERATION_REDUCTION_XDESC,
                                                       CUDNN_TYPE_BACKEND_DESCRIPTOR,
                                                       1,
                                                       &x_backend_desc));

        // 4. Set the output tensor (Y) descriptor attribute
        auto y_backend_desc = tensors.at(Y->second->get_uid())->get_raw_desc();

        _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(reduction_operation.get_raw_desc(),
                                                       CUDNN_ATTR_OPERATION_REDUCTION_YDESC,
                                                       CUDNN_TYPE_BACKEND_DESCRIPTOR,
                                                       1,
                                                       &y_backend_desc));

        // 5. Set the optional group offset tensor descriptor
        if (Group_offset != attributes.inputs.end() && Group_offset->second != nullptr) {
#if (CUDNN_VERSION >= 92400)
            RETURN_CUDNN_FRONTEND_ERROR_IF(detail::get_backend_version() < 92400,
                                           error_code_t::GRAPH_NOT_SUPPORTED,
                                           "Reduction group_offset is not supported in cudnn version < 9.24.0");
            auto group_offset_backend_desc = tensors.at(Group_offset->second->get_uid())->get_raw_desc();

            _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(reduction_operation.get_raw_desc(),
                                                           CUDNN_ATTR_OPERATION_REDUCTION_GROUP_OFFSET_DESC,
                                                           CUDNN_TYPE_BACKEND_DESCRIPTOR,
                                                           1,
                                                           &group_offset_backend_desc));
#else
            return {error_code_t::GRAPH_NOT_SUPPORTED,
                    "Reduction group_offset is not supported in cudnn version < 9.24.0"};
#endif
        }

        // 6. Finalize the operation descriptor

        _CUDNN_CHECK_CUDNN_ERROR(detail::finalize(reduction_operation.get_raw_desc()));

        operations.push_back(std::make_shared<Operation_v8>(std::move(reduction_operation)));

        auto const& non_virtual_uids = attributes.get_non_virtual_uids();
        uids_involved_in_operations.insert(non_virtual_uids.begin(), non_virtual_uids.end());
        return {error_code_t::OK, ""};
    }

#ifndef CUDNN_FRONTEND_SKIP_JSON_LIB
    virtual void
    serialize(json& j) const override final {
        j = attributes;
        j.update(R"({"tag": "REDUCTION"})"_json);
    }
#endif
};

inline void
INode::reduction(std::shared_ptr<Tensor_attributes> a,
                 Reduction_attributes attributes,
                 std::shared_ptr<Tensor_attributes> c) {
    attributes.inputs[Reduction_attributes::input_names::X]   = a;
    attributes.outputs[Reduction_attributes::output_names::Y] = c;
    sub_nodes.emplace_back(std::make_unique<ReductionNode>(std::move(attributes), context));
}

inline void
INode::reduction(std::shared_ptr<Tensor_attributes> a,
                 std::shared_ptr<Tensor_attributes> group_offset,
                 Reduction_attributes attributes,
                 std::shared_ptr<Tensor_attributes> c) {
    attributes.inputs[Reduction_attributes::input_names::X]            = a;
    attributes.inputs[Reduction_attributes::input_names::Group_offset] = group_offset;
    attributes.outputs[Reduction_attributes::output_names::Y]          = c;
    sub_nodes.emplace_back(std::make_unique<ReductionNode>(std::move(attributes), context));
}

inline std::shared_ptr<Tensor_attributes>
INode::reduction(std::shared_ptr<Tensor_attributes> input, Reduction_attributes attributes) {
    attributes.inputs[Reduction_attributes::input_names::X] = input;
    auto Y = attributes.outputs[Reduction_attributes::output_names::Y] = output_tensor(attributes.name + "::Y");

    sub_nodes.emplace_back(std::make_unique<ReductionNode>(std::move(attributes), context));
    return Y;
}

inline std::shared_ptr<Tensor_attributes>
INode::reduction(std::shared_ptr<Tensor_attributes> input,
                 std::shared_ptr<Tensor_attributes> group_offset,
                 Reduction_attributes attributes) {
    attributes.inputs[Reduction_attributes::input_names::X]            = input;
    attributes.inputs[Reduction_attributes::input_names::Group_offset] = group_offset;
    auto Y = attributes.outputs[Reduction_attributes::output_names::Y] = output_tensor(attributes.name + "::Y");

    sub_nodes.emplace_back(std::make_unique<ReductionNode>(std::move(attributes), context));
    return Y;
}
}  // namespace cudnn_frontend::graph